From ee7203281b5dceb3158f057299ce7e56bfaba761 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 21 Aug 2026 12:39:57 +0000 Subject: [PATCH 01/44] fix(ptu): take the router as an argument instead of the proxy module global The rollup read litellm.proxy.proxy_server.llm_router out of sys.modules, so a run priced and swept whatever deployments anything else in the process had left on that module. Under xdist the shard's module-to-worker assignment varies per run, which made three rollup tests fail or pass on the same commit depending on ordering. Callers now hand the router in, and the proxy's scheduled job passes its own. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/proxy/proxy_server.py | 1 + .../spend_tracking/ptu_flat_cost_rollup.py | 40 ++-- .../test_ptu_flat_cost_rollup.py | 206 +++++++++--------- tests/test_litellm/proxy/test_proxy_server.py | 29 +++ 4 files changed, 147 insertions(+), 129 deletions(-) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 9ee62f94647..2e57d3e0708 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -9089,6 +9089,7 @@ class ProxyStartupEvent: prisma_client, pod_lock_manager=proxy_logging_obj.db_spend_update_writer.pod_lock_manager, alert=_alert_ptu_rollup_failure, + router=llm_router, ) scheduler.add_job( diff --git a/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py b/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py index f1f7248c064..a4eac992d89 100644 --- a/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py +++ b/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py @@ -14,7 +14,6 @@ and share the existing unique constraint. import asyncio import json -import sys from collections.abc import Awaitable, Callable, Mapping from dataclasses import dataclass from datetime import date, datetime, time, timedelta, timezone @@ -327,16 +326,6 @@ class _LoadedDeployments: config_sourced: bool -def _running_router() -> object | None: - """The proxy's router, or None outside a running proxy. - - Read out of ``sys.modules`` rather than imported, so a rollup driven from a test or a - script does not pull the whole proxy server in behind it. - """ - proxy_server: Final = sys.modules.get("litellm.proxy.proxy_server") - return getattr(proxy_server, "llm_router", None) if proxy_server is not None else None - - def _config_deployments(router: object | None, *, owned_by_db: frozenset[str]) -> tuple[_PTUDeployment, ...]: """Deployments the router holds that no ``LiteLLM_ProxyModelTable`` row owns. @@ -357,15 +346,17 @@ def _config_deployments(router: object | None, *, owned_by_db: frozenset[str]) - ) -async def _load_ptu_models(prisma_client: "PrismaClient") -> _LoadedDeployments: +async def _load_ptu_models(prisma_client: "PrismaClient", *, router: object | None) -> _LoadedDeployments: """Every deployment carrying valid manual PTU config, and every id the scan saw. Reserved capacity is billed by the provider whichever file declared it, so a deployment the proxy only knows from config.yaml accrues alongside the stored ones. + The router is handed in rather than read off the proxy module, so a run prices exactly + the deployments its caller declares and nothing a co-resident process left behind. """ rows: Final = await prisma_client.db.litellm_proxymodeltable.find_many() db_ids: Final = frozenset(model_id for row in rows if (model_id := str(getattr(row, "model_id", "") or ""))) - config_records: Final = _config_deployments(_running_router(), owned_by_db=db_ids) + config_records: Final = _config_deployments(router, owned_by_db=db_ids) models: Final = tuple( parsed for parsed in (_parse_ptu_model(row) for row in (*rows, *config_records)) if parsed is not None ) @@ -382,6 +373,7 @@ async def run_ptu_flat_cost_rollup( prisma_client: "PrismaClient", target_date: date | None = None, may_prune: bool = True, + router: object | None = None, ) -> RollupResult: """Rollup one UTC day of flat PTU cost across all PTU-configured model deployments. @@ -406,7 +398,7 @@ async def run_ptu_flat_cost_rollup( date_str: Final = day.isoformat() run_started: Final = datetime.now(timezone.utc) - loaded: Final = await _load_ptu_models(prisma_client) + loaded: Final = await _load_ptu_models(prisma_client, router=router) ptu_models: Final = loaded.models charges: Final = _aggregate_charges(ptu_models, day) @@ -527,6 +519,7 @@ async def _existing_sentinel_keys( async def run_ptu_flat_cost_backfill( prisma_client: "PrismaClient", today: date | None = None, + router: object | None = None, ) -> BackfillResult: """Price the elapsed days of every PTU window that carry no sentinel row yet. @@ -546,7 +539,7 @@ async def run_ptu_flat_cost_backfill( verbose_proxy_logger.warning("PTU backfill: prisma_client is None, skipping") return BackfillResult(start=end, end=end, days_scanned=0, rows_written=0) - ptu_models: Final = (await _load_ptu_models(prisma_client)).models + ptu_models: Final = (await _load_ptu_models(prisma_client, router=router)).models days: Final = _backfill_window(ptu_models, end) if not days: @@ -591,6 +584,7 @@ async def run_scheduled_ptu_rollup( pod_lock_manager: "PodLockManager | None" = None, target_date: date | None = None, alert: Callable[[str], Awaitable[None]] | None = None, + router: object | None = None, ) -> RollupResult | None: """Run the daily rollup under a cross-pod lock so only one proxy reconciles a day. @@ -615,7 +609,7 @@ async def run_scheduled_ptu_rollup( return None if pod_lock_manager is None or pod_lock_manager.redis_cache is None: - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False, router=router) if not await pod_lock_manager.acquire_lock(cronjob_id=PTU_ROLLUP_JOB_ID, ttl=PTU_ROLLUP_LOCK_TTL_SECONDS): if await _lock_is_held(pod_lock_manager): @@ -629,10 +623,10 @@ async def run_scheduled_ptu_rollup( "PTU rollup: could not take the rollup lock and no other pod holds it, " "running unguarded rather than skipping the day" ) - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False, router=router) try: - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=True) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=True, router=router) finally: await pod_lock_manager.release_lock(cronjob_id=PTU_ROLLUP_JOB_ID) @@ -657,6 +651,7 @@ async def _run_and_alert( target_date: date | None, alert: "Callable[[str], Awaitable[None]] | None", may_prune: bool = True, + router: object | None = None, ) -> RollupResult: """Reconcile the day, catch up any days left unpriced, and alert on charges that did not land. @@ -669,7 +664,9 @@ async def _run_and_alert( explicit date means reconcile exactly that day, so it stays a single-day operation. Its failure is contained: the day's own result is returned either way. """ - result: Final = await run_ptu_flat_cost_rollup(prisma_client, target_date=target_date, may_prune=may_prune) + result: Final = await run_ptu_flat_cost_rollup( + prisma_client, target_date=target_date, may_prune=may_prune, router=router + ) if result.rows_failed: await _deliver_alert( alert, @@ -686,7 +683,7 @@ async def _run_and_alert( "by the provider with nothing attributing it here. Extend the window, or retire the deployment.", ) if target_date is None: - await _backfill_and_alert(prisma_client, alert=alert) + await _backfill_and_alert(prisma_client, alert=alert, router=router) return result @@ -694,6 +691,7 @@ async def _backfill_and_alert( prisma_client: "PrismaClient", *, alert: "Callable[[str], Awaitable[None]] | None", + router: object | None = None, ) -> None: """Catch up unpriced PTU days, alerting on charges that did not land. @@ -701,7 +699,7 @@ async def _backfill_and_alert( caller whatever the catch-up pass does. """ try: - backfill: Final = await run_ptu_flat_cost_backfill(prisma_client) + backfill: Final = await run_ptu_flat_cost_backfill(prisma_client, router=router) except Exception as exc: # noqa: BLE001 # the catch-up pass must not fail the day's rollup verbose_proxy_logger.error("PTU backfill: catch-up pass failed, the day's rollup still stands: %s", exc) return diff --git a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py index e039455607d..17a487ddd4b 100644 --- a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py +++ b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py @@ -1767,16 +1767,20 @@ async def test_the_prune_cutoff_allows_for_clock_skew_between_hosts(): @pytest.mark.asyncio -async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(monkeypatch): +async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(): """Staleness alone stops being evidence once two hosts hold different configuration: a row this run never considered belongs to a deployment another host is pricing from its own file, and sweeping it drops that charge.""" table = _FakeSentinelTable() table.seed("t", DAY, "dep-elsewhere", 480.0, updated_at=datetime(2020, 1, 1, tzinfo=timezone.utc)) entry = _router_entry(model_id="cfg-here", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY) + await run_scheduled_ptu_rollup( + _prisma_for([], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(entry), + ) assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "dep-elsewhere") in table.rows assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "cfg-here") in table.rows @@ -1784,7 +1788,7 @@ async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(monkeypat @pytest.mark.asyncio -async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged(monkeypatch): +async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged(): """The accepted cost of bounding the prune, driven through the sequence that produces it: charge the day while the deployment exists, remove it, run the day again. Nothing scans it now, so nothing may judge its row, and the amount it was billed stands.""" @@ -1793,18 +1797,19 @@ async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged( live_row = _model_row(model_id="dep-live", model_info=ptu) doomed_row = _model_row(model_id="dep-doomed", model_info=ptu) charged_key = ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "dep-doomed") - monkeypatch.setattr( - ptu_rollup, "_running_router", lambda: _router_holding(_router_entry(model_id="cfg", model_info=dict(ptu))) - ) + router = _router_holding(_router_entry(model_id="cfg", model_info=dict(ptu))) await run_scheduled_ptu_rollup( - _prisma_for([live_row, doomed_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for([live_row, doomed_row], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=router, ) billed = table.rows[charged_key]["ptu_flat_cost"] table.rows[charged_key]["updated_at"] = datetime(2020, 1, 1, tzinfo=timezone.utc) await run_scheduled_ptu_rollup( - _prisma_for([live_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for([live_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY, router=router ) assert table.rows[charged_key]["ptu_flat_cost"] == billed @@ -1842,7 +1847,7 @@ async def test_every_deployment_that_prices_is_inside_the_set_that_bounds_the_pr table, ) - loaded = await ptu_rollup._load_ptu_models(prisma) + loaded = await ptu_rollup._load_ptu_models(prisma, router=None) assert {model.model_id for model in loaded.models} <= loaded.scanned_ids assert loaded.scanned_ids == {"dep-a", "dep-b", "dep-unpriced"} @@ -1858,7 +1863,7 @@ async def test_a_priced_deployment_is_in_the_bound_even_with_an_id_the_scan_skip _FakeSentinelTable(), ) - loaded = await ptu_rollup._load_ptu_models(prisma) + loaded = await ptu_rollup._load_ptu_models(prisma, router=None) assert {model.model_id for model in loaded.models} <= loaded.scanned_ids @@ -1872,13 +1877,13 @@ async def test_the_prune_splits_the_id_set_across_statements(monkeypatch): table = _FakeSentinelTable() ptu = {"ptu_count": 5, "cost_per_ptu_per_hour": 2.0, "team_id": "t"} deployments = [_model_row(model_id=f"dep-{n}", model_info=ptu) for n in range(4)] - monkeypatch.setattr( - ptu_rollup, "_running_router", lambda: _router_holding(_router_entry(model_id="dep-4", model_info=dict(ptu))) - ) table.seed("t", DAY, "dep-3", 480.0, updated_at=datetime(2020, 1, 1, tzinfo=timezone.utc)) await run_scheduled_ptu_rollup( - _prisma_for(deployments, table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for(deployments, table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(_router_entry(model_id="dep-4", model_info=dict(ptu))), ) chunks = [call["model"]["in"] for call in table.delete_many_calls] @@ -1911,140 +1916,124 @@ def _router_holding(*entries): @pytest.mark.asyncio -async def test_a_config_declared_deployment_is_priced(monkeypatch): +async def test_a_config_declared_deployment_is_priced(): """The whole point. A PTU deployment the proxy only knows from config.yaml is not in LiteLLM_ProxyModelTable, so a DB-only scan bills the provider's reservation to nobody.""" entry = _router_entry(model_id="cfg-1", model_name="gpt-4o-ptu", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(entry)) assert [(m.model_id, m.model_name, m.team_id) for m in loaded.models] == [("cfg-1", "gpt-4o-ptu", "t")] assert "cfg-1" in loaded.scanned_ids @pytest.mark.asyncio -async def test_a_database_backed_router_entry_is_not_counted_twice(monkeypatch): +async def test_a_database_backed_router_entry_is_not_counted_twice(): """Every deployment loaded from the table is also in the router, flagged db_model. Pricing both copies would write two charges for one reservation.""" row = _model_row(model_id="db-1", model_info=dict(_VALID_PTU)) mirrored = _router_entry(model_id="db-1", model_info={**_VALID_PTU, "db_model": True}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(mirrored)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([row], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["db-1"] - - -@pytest.mark.asyncio -async def test_a_router_entry_sharing_an_id_with_the_table_is_priced_once(monkeypatch): - """db_model is data the router carries rather than something this module controls, so the - id anti-join is what actually maps onto the failure: two charges under one id.""" - row = _model_row(model_id="both-1", model_info=dict(_VALID_PTU)) - unflagged = _router_entry(model_id="both-1", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(unflagged)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([row], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["both-1"] - - -@pytest.mark.asyncio -async def test_a_client_credential_clone_is_not_priced(monkeypatch): - """Supplying an api_key on a request mints a clone of the deployment under a fresh id, - carrying the source's PTU config. Pricing it bills one reservation per distinct caller key.""" - source = _router_entry(model_id="cfg-1", model_info=dict(_VALID_PTU)) - clone = _router_entry(model_id="cfg-1-clone", model_info={**_VALID_PTU, "original_model_id": "cfg-1"}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(source, clone)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["cfg-1"] - - -@pytest.mark.asyncio -async def test_a_config_deployment_without_ptu_config_is_scanned_but_not_priced(monkeypatch): - """It has to stay in the scanned set or its leftover sentinel rows become unprunable.""" - entry = _router_entry(model_id="cfg-plain", model_info={"team_id": "t"}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) - - assert loaded.models == () - assert "cfg-plain" in loaded.scanned_ids - - -@pytest.mark.asyncio -async def test_no_router_in_the_process_prices_the_database_alone(monkeypatch): - """The rollup is importable and callable outside a running proxy.""" - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: None) loaded = await ptu_rollup._load_ptu_models( - _prisma_for([_model_row(model_id="db-1", model_info=dict(_VALID_PTU))], _FakeSentinelTable()) + _prisma_for([row], _FakeSentinelTable()), router=_router_holding(mirrored) ) assert [m.model_id for m in loaded.models] == ["db-1"] @pytest.mark.asyncio -async def test_a_config_deployment_is_charged_end_to_end(monkeypatch): +async def test_a_router_entry_sharing_an_id_with_the_table_is_priced_once(): + """db_model is data the router carries rather than something this module controls, so the + id anti-join is what actually maps onto the failure: two charges under one id.""" + row = _model_row(model_id="both-1", model_info=dict(_VALID_PTU)) + unflagged = _router_entry(model_id="both-1", model_info=dict(_VALID_PTU)) + + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([row], _FakeSentinelTable()), router=_router_holding(unflagged) + ) + + assert [m.model_id for m in loaded.models] == ["both-1"] + + +@pytest.mark.asyncio +async def test_a_client_credential_clone_is_not_priced(): + """Supplying an api_key on a request mints a clone of the deployment under a fresh id, + carrying the source's PTU config. Pricing it bills one reservation per distinct caller key.""" + source = _router_entry(model_id="cfg-1", model_info=dict(_VALID_PTU)) + clone = _router_entry(model_id="cfg-1-clone", model_info={**_VALID_PTU, "original_model_id": "cfg-1"}) + + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([], _FakeSentinelTable()), router=_router_holding(source, clone) + ) + + assert [m.model_id for m in loaded.models] == ["cfg-1"] + + +@pytest.mark.asyncio +async def test_a_config_deployment_without_ptu_config_is_scanned_but_not_priced(): + """It has to stay in the scanned set or its leftover sentinel rows become unprunable.""" + entry = _router_entry(model_id="cfg-plain", model_info={"team_id": "t"}) + + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(entry)) + + assert loaded.models == () + assert "cfg-plain" in loaded.scanned_ids + + +@pytest.mark.asyncio +async def test_no_router_in_the_process_prices_the_database_alone(): + """The rollup is importable and callable outside a running proxy.""" + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([_model_row(model_id="db-1", model_info=dict(_VALID_PTU))], _FakeSentinelTable()), router=None + ) + + assert [m.model_id for m in loaded.models] == ["db-1"] + + +@pytest.mark.asyncio +async def test_a_config_deployment_is_charged_end_to_end(): """Through the scheduled entry point, so the charge lands in a sentinel row rather than stopping at the loader.""" table = _FakeSentinelTable() entry = _router_entry(model_id="cfg-1", model_name="gpt-4o-ptu", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY) + await run_scheduled_ptu_rollup( + _prisma_for([], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(entry), + ) assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "cfg-1") in table.rows @pytest.mark.asyncio -async def test_a_stale_database_backed_router_entry_is_not_treated_as_config(monkeypatch): +async def test_a_stale_database_backed_router_entry_is_not_treated_as_config(): """The reconcile can leave a deployment on the router after its row is gone. The id anti-join cannot see that one, so the flag is what keeps it from being priced as though config.yaml had declared it.""" stale = _router_entry(model_id="db-gone", model_info={**_VALID_PTU, "db_model": True}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(stale)) - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(stale)) assert loaded.models == () -def test_the_router_lookup_reads_the_proxys_own_global(): - """Every other config test replaces this helper, so without one test driving the real - body a typo in the module path or the attribute name leaves the whole feature dead in - production with the suite still green.""" - import sys - import types as _types +@pytest.mark.asyncio +async def test_a_router_left_on_the_proxy_module_is_not_scanned(monkeypatch): + """A run scans the router its caller hands it and nothing else. Reading the proxy module's + global instead made every run depend on whatever else in the process had set one, which + is what a caller passing no router is asking not to happen.""" + import litellm.proxy.proxy_server as proxy_server - assert ptu_rollup._running_router() is None or "litellm.proxy.proxy_server" in sys.modules + ambient = _router_holding(_router_entry(model_id="ambient-1", model_info=dict(_VALID_PTU))) + monkeypatch.setattr(proxy_server, "llm_router", ambient, raising=False) - sentinel = object() - stub = _types.SimpleNamespace(llm_router=sentinel) - real = sys.modules.get("litellm.proxy.proxy_server") - sys.modules["litellm.proxy.proxy_server"] = stub - try: - assert ptu_rollup._running_router() is sentinel - del stub.llm_router - assert ptu_rollup._running_router() is None - finally: - if real is None: - del sys.modules["litellm.proxy.proxy_server"] - else: - sys.modules["litellm.proxy.proxy_server"] = real + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=None) - -def test_the_router_lookup_returns_none_outside_a_proxy(): - import sys - - real = sys.modules.pop("litellm.proxy.proxy_server", None) - try: - assert ptu_rollup._running_router() is None - finally: - if real is not None: - sys.modules["litellm.proxy.proxy_server"] = real + assert loaded.models == () + assert loaded.scanned_ids == frozenset() + assert loaded.config_sourced is False @pytest.mark.parametrize("chunk", [None, ("dep-a", "dep-b")], ids=["unbounded", "bounded"]) @@ -2063,7 +2052,7 @@ def test_the_prune_filter_is_a_plain_dict(chunk): @pytest.mark.asyncio -async def test_the_catch_up_pass_reaches_a_config_declared_deployment(monkeypatch): +async def test_the_catch_up_pass_reaches_a_config_declared_deployment(): """The catch-up shares the loader, so config deployments join it without being wired in. That is what prices the elapsed days of a reservation declared before today.""" table = _FakeSentinelTable() @@ -2073,9 +2062,10 @@ async def test_the_catch_up_pass_reaches_a_config_declared_deployment(monkeypatc model_id="cfg-back", model_info={"ptu_count": 100, "cost_per_ptu_per_hour": 0.02, "team_id": "t", "ptu_effective_from": started}, ) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True)) + await run_scheduled_ptu_rollup( + _prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), router=_router_holding(entry) + ) charged = sorted(day for (_, day, _, model) in table.rows if model == "cfg-back") yesterday = (now.date() - timedelta(days=1)).isoformat() diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 83e9095c8ec..71fb184eb4b 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -11048,6 +11048,35 @@ async def test_ptu_rollup_job_registered_at_startup(monkeypatch): assert scheduler.get_job(PTU_ROLLUP_JOB_ID) is not None +@pytest.mark.asyncio +async def test_ptu_rollup_job_hands_the_rollup_the_proxys_router(monkeypatch): + """The rollup prices PTU deployments declared in config.yaml, which only the router + knows about. It takes the router as an argument, so nothing but this call site puts the + proxy's own router in front of it: without it that half of the feature is dead.""" + monkeypatch.delenv("STORE_MODEL_IN_DB", raising=False) + from litellm.proxy.spend_tracking import ptu_flat_cost_rollup + from litellm.proxy.spend_tracking.ptu_feature_flag import PTU_COST_ATTRIBUTION_ENV_VAR + from litellm.proxy.spend_tracking.ptu_flat_cost_rollup import PTU_ROLLUP_JOB_ID + + monkeypatch.setenv(PTU_COST_ATTRIBUTION_ENV_VAR, "true") + calls = [] + monkeypatch.setattr( + ptu_flat_cost_rollup, + "run_scheduled_ptu_rollup", + AsyncMock(side_effect=lambda *args, **kwargs: calls.append(kwargs)), + ) + + scheduler = await _run_scheduled_background_jobs() + + import litellm.proxy.proxy_server as ps + + router = MagicMock() + monkeypatch.setattr(ps, "llm_router", router) + await scheduler.get_job(PTU_ROLLUP_JOB_ID).func() + + assert [call["router"] for call in calls] == [router] + + @pytest.mark.asyncio async def test_ptu_rollup_job_not_registered_without_opt_in(monkeypatch): """Without LITELLM_ENABLE_PTU_COST_ATTRIBUTION the rollup never runs, so no sentinel row From 0b938e37f46d1149db2e99198b1036d207aeb49a Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 19 Aug 2026 09:52:33 +0000 Subject: [PATCH 02/44] test: invalidate memoized model-cost lookups between unit tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/conftest.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index 1fe73b552da..62c95cb100b 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -375,6 +375,9 @@ def isolate_litellm_state(): litellm.in_memory_llm_clients_cache.flush_cache() image_handling_module.in_memory_cache.flush_cache() _reset_module_level_aws_auth_caches() + # litellm.get_model_info() memoizes ModelInfo built from litellm.model_cost, so a + # test that rebinds the cost map leaves later tests pricing against the old map. + litellm_utils_module._invalidate_model_cost_lowercase_map() # Clear all callback lists to prevent cross-test contamination if hasattr(litellm, "callbacks"): @@ -418,6 +421,7 @@ def isolate_litellm_state(): litellm_utils_module._runtime_registered_model_cost.clear() litellm_utils_module._runtime_registered_model_cost.update(original_runtime_registered_model_cost) + litellm_utils_module._invalidate_model_cost_lowercase_map() for _router in tuple(litellm_router_module._live_routers): litellm_router_module._live_routers.discard(_router) From e4a72c587d8dfb372185fd9f6ec9dd8cf2ead111 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 20 Aug 2026 09:49:03 +0000 Subject: [PATCH 03/44] fix(ci): retry transient PyPI license lookups Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/code_coverage_tests/check_licenses.py | 66 ++++++++++++++----- tests/test_litellm/test_check_licenses.py | 71 +++++++++++++++++++++ 2 files changed, 120 insertions(+), 17 deletions(-) diff --git a/tests/code_coverage_tests/check_licenses.py b/tests/code_coverage_tests/check_licenses.py index 389e534b1ff..67d1d91a6f7 100644 --- a/tests/code_coverage_tests/check_licenses.py +++ b/tests/code_coverage_tests/check_licenses.py @@ -5,8 +5,9 @@ import json from pathlib import Path import re import sys +import time import tomllib -from typing import Dict, List, Optional, Set, Tuple +from typing import Callable, Dict, Final, List, Optional, Set, Tuple from packaging.requirements import Requirement import requests @@ -37,6 +38,8 @@ DEFAULT_TRANSITIVE_PIN_PACKAGES = ( # of the identifier, not an operator. _SPDX_OPERATOR_SPLIT = re.compile(r"\s+(?:OR|AND)\s+") _SPDX_WITH_SUFFIX = re.compile(r"\s+WITH\s+.*", re.DOTALL) +_PYPI_FETCH_ATTEMPTS: Final[int] = 3 +_PYPI_FETCH_BACKOFF_SECONDS: Final[float] = 0.5 @dataclass @@ -50,7 +53,10 @@ class PackageLicense: class LicenseChecker: def __init__( - self, config_file: Path = Path("./tests/code_coverage_tests/liccheck.ini") + self, + config_file: Path = Path("./tests/code_coverage_tests/liccheck.ini"), + http_get: Optional[Callable[..., requests.Response]] = None, + sleep: Optional[Callable[[float], None]] = None, ): if not config_file.exists(): print(f"Error: Config file {config_file} not found") @@ -79,6 +85,8 @@ class LicenseChecker: # Track package results self.package_results: List[PackageLicense] = [] + self._http_get = http_get + self._sleep = sleep @staticmethod def _normalize_package_name(package_name: str) -> str: @@ -123,21 +131,45 @@ class LicenseChecker: last resort derives the license from the ``License :: OSI Approved :: ...`` trove classifiers. """ - try: - url = f"https://pypi.org/pypi/{package_name}/{version}/json" - response = requests.get(url, timeout=10) - response.raise_for_status() - info = response.json().get("info", {}) or {} - return ( - info.get("license_expression") - or info.get("license") - or self._license_from_classifiers(info.get("classifiers") or []) - ) - except Exception as e: - print( - f"Warning: Failed to fetch license for {package_name} {version}: {str(e)}" - ) - return None + url = f"https://pypi.org/pypi/{package_name}/{version}/json" + http_get = self._http_get if self._http_get is not None else requests.get + sleep = self._sleep if self._sleep is not None else time.sleep + + for attempt in range(_PYPI_FETCH_ATTEMPTS): + try: + response = http_get(url, timeout=10) + response.raise_for_status() + info = response.json().get("info", {}) or {} + return ( + info.get("license_expression") + or info.get("license") + or self._license_from_classifiers(info.get("classifiers") or []) + ) + except requests.HTTPError as error: + status_code = error.response.status_code if error.response is not None else None + if status_code != 429 and (status_code is None or status_code < 500): + print( + f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" + ) + return None + if attempt == _PYPI_FETCH_ATTEMPTS - 1: + print( + f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" + ) + return None + sleep(_PYPI_FETCH_BACKOFF_SECONDS) + except (requests.ConnectionError, requests.Timeout) as error: + if attempt == _PYPI_FETCH_ATTEMPTS - 1: + print( + f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" + ) + return None + sleep(_PYPI_FETCH_BACKOFF_SECONDS) + except Exception as error: + print( + f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" + ) + return None @staticmethod def _license_from_classifiers(classifiers: List[str]) -> Optional[str]: diff --git a/tests/test_litellm/test_check_licenses.py b/tests/test_litellm/test_check_licenses.py index 4d72f185a25..1218e44fade 100644 --- a/tests/test_litellm/test_check_licenses.py +++ b/tests/test_litellm/test_check_licenses.py @@ -12,6 +12,8 @@ import os import sys from pathlib import Path +import requests + _CODE_COVERAGE_DIR = os.path.join( os.path.dirname(os.path.abspath(__file__)), "..", "code_coverage_tests" ) @@ -122,6 +124,75 @@ def test_get_license_returns_none_on_request_failure(monkeypatch): assert checker.get_package_license_from_pypi("pkg", "1.0.0") is None +def test_get_license_retries_connection_error_then_resolves_license(): + responses = iter( + ( + requests.ConnectionError("connection reset"), + requests.ConnectionError("connection reset"), + _FakeResponse({"info": {"license_expression": "MIT"}}), + ) + ) + calls = [] + sleeps = [] + + def _fake_get(url, timeout=None): + calls.append((url, timeout)) + response = next(responses) + if isinstance(response, Exception): + raise response + return response + + checker = check_licenses.LicenseChecker( + config_file=_LICCHECK_INI, + http_get=_fake_get, + sleep=sleeps.append, + ) + + assert checker.get_package_license_from_pypi("pkg", "1.0.0") == "MIT" + assert len(calls) == 3 + assert len(sleeps) == 2 + + +def test_get_license_does_not_retry_not_found_http_error(): + response = requests.Response() + response.status_code = 404 + calls = [] + sleeps = [] + + def _fake_get(url, timeout=None): + calls.append((url, timeout)) + raise requests.HTTPError("not found", response=response) + + checker = check_licenses.LicenseChecker( + config_file=_LICCHECK_INI, + http_get=_fake_get, + sleep=sleeps.append, + ) + + assert checker.get_package_license_from_pypi("pkg", "1.0.0") is None + assert len(calls) == 1 + assert sleeps == [] + + +def test_get_license_returns_none_after_connection_retry_limit(): + calls = [] + sleeps = [] + + def _fake_get(url, timeout=None): + calls.append((url, timeout)) + raise requests.ConnectionError("connection reset") + + checker = check_licenses.LicenseChecker( + config_file=_LICCHECK_INI, + http_get=_fake_get, + sleep=sleeps.append, + ) + + assert checker.get_package_license_from_pypi("pkg", "1.0.0") is None + assert len(calls) == 3 + assert len(sleeps) == 2 + + # -------------------------------------------------------------------------- # is_license_acceptable: SPDX identifiers and compound expressions # -------------------------------------------------------------------------- From 134b6252e0012e92ac59c2f335af354b941aba4e Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 20 Aug 2026 09:54:38 +0000 Subject: [PATCH 04/44] refactor(ci): simplify PyPI license retries Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/code_coverage_tests/check_licenses.py | 42 ++++++++++----------- 1 file changed, 20 insertions(+), 22 deletions(-) diff --git a/tests/code_coverage_tests/check_licenses.py b/tests/code_coverage_tests/check_licenses.py index 67d1d91a6f7..158e25180e1 100644 --- a/tests/code_coverage_tests/check_licenses.py +++ b/tests/code_coverage_tests/check_licenses.py @@ -7,7 +7,7 @@ import re import sys import time import tomllib -from typing import Callable, Dict, Final, List, Optional, Set, Tuple +from typing import Callable, Dict, Final, List, Optional, Protocol, Set, Tuple from packaging.requirements import Requirement import requests @@ -42,6 +42,11 @@ _PYPI_FETCH_ATTEMPTS: Final[int] = 3 _PYPI_FETCH_BACKOFF_SECONDS: Final[float] = 0.5 +class _HttpGet(Protocol): + def __call__(self, url: str, *, timeout: float) -> requests.Response: + ... + + @dataclass class PackageLicense: name: str @@ -55,7 +60,7 @@ class LicenseChecker: def __init__( self, config_file: Path = Path("./tests/code_coverage_tests/liccheck.ini"), - http_get: Optional[Callable[..., requests.Response]] = None, + http_get: Optional[_HttpGet] = None, sleep: Optional[Callable[[float], None]] = None, ): if not config_file.exists(): @@ -145,31 +150,24 @@ class LicenseChecker: or info.get("license") or self._license_from_classifiers(info.get("classifiers") or []) ) - except requests.HTTPError as error: - status_code = error.response.status_code if error.response is not None else None - if status_code != 429 and (status_code is None or status_code < 500): - print( - f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" - ) - return None - if attempt == _PYPI_FETCH_ATTEMPTS - 1: - print( - f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" - ) - return None - sleep(_PYPI_FETCH_BACKOFF_SECONDS) - except (requests.ConnectionError, requests.Timeout) as error: - if attempt == _PYPI_FETCH_ATTEMPTS - 1: - print( - f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" - ) - return None - sleep(_PYPI_FETCH_BACKOFF_SECONDS) except Exception as error: + if self._is_retryable_pypi_error(error) and attempt < _PYPI_FETCH_ATTEMPTS - 1: + sleep(_PYPI_FETCH_BACKOFF_SECONDS) + continue print( f"Warning: Failed to fetch license for {package_name} {version}: {str(error)}" ) return None + return None + + @staticmethod + def _is_retryable_pypi_error(error: Exception) -> bool: + if isinstance(error, (requests.ConnectionError, requests.Timeout)): + return True + if not isinstance(error, requests.HTTPError) or error.response is None: + return False + status_code = error.response.status_code + return status_code == 429 or status_code >= 500 @staticmethod def _license_from_classifiers(classifiers: List[str]) -> Optional[str]: From 8deade4f345bcd983c721f0862fcd8ad30dfebda Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sun, 23 Aug 2026 10:02:41 +0000 Subject: [PATCH 05/44] test(ptu): drop the assertion on the flag removed upstream Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../proxy/spend_tracking/test_ptu_flat_cost_rollup.py | 1 - 1 file changed, 1 deletion(-) diff --git a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py index 76fa41c83be..8f25cffecf5 100644 --- a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py +++ b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py @@ -2034,7 +2034,6 @@ async def test_a_router_left_on_the_proxy_module_is_not_scanned(monkeypatch): assert loaded.models == () assert loaded.scanned_ids == frozenset() - assert loaded.config_sourced is False def test_the_prune_filter_is_a_plain_dict(): From 54ea379c91ab4c60e45c4b671fdd76003fea6188 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 25 Aug 2026 10:03:57 +0000 Subject: [PATCH 06/44] fix(tests): drain the logging worker queue between MCP tests Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/mcp_tests/conftest.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/tests/mcp_tests/conftest.py b/tests/mcp_tests/conftest.py index d1dc3ec7216..e46c03b5498 100644 --- a/tests/mcp_tests/conftest.py +++ b/tests/mcp_tests/conftest.py @@ -42,6 +42,22 @@ def setup_and_teardown(): asyncio.set_event_loop(None) # Remove the reference to the loop +@pytest.fixture(scope="function", autouse=True) +async def drain_logging_worker(): + """ + The logging queue is bound to the running loop, so anything left queued when a test's loop + goes away is carried onto the next test's loop and fires against its callbacks. + """ + from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER + + yield + + try: + await asyncio.wait_for(GLOBAL_LOGGING_WORKER.clear_queue(), timeout=10) + except asyncio.TimeoutError: + pass + + def pytest_collection_modifyitems(config, items): # Separate tests in 'test_amazing_proxy_custom_logger.py' and other tests custom_logger_tests = [ From ab160fb9537dee34183e6a5cb730791b7e1791a0 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 25 Aug 2026 13:26:48 +0000 Subject: [PATCH 07/44] fix(model_prices): sync gpt-5.6-sol bedrock rates, add gpt-5.6-cyber, fix claude 3 1h cache writes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 80 +++++++++++++------ model_prices_and_context_window.json | 80 +++++++++++++------ ..._cross_region_inference_profile_mapping.py | 24 +++--- ...bedrock_mantle_responses_transformation.py | 13 ++- 4 files changed, 131 insertions(+), 66 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 9f953e11df1..73e16715cf0 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -12282,7 +12282,7 @@ }, "claude-3-haiku-20240307": { "cache_creation_input_token_cost": 3e-07, - "cache_creation_input_token_cost_above_1hr": 6e-06, + "cache_creation_input_token_cost_above_1hr": 5e-07, "cache_read_input_token_cost": 3e-08, "deprecation_date": "2026-04-20", "input_cost_per_token": 2.5e-07, @@ -12301,7 +12301,7 @@ }, "claude-3-opus-20240229": { "cache_creation_input_token_cost": 1.875e-05, - "cache_creation_input_token_cost_above_1hr": 6e-06, + "cache_creation_input_token_cost_above_1hr": 3e-05, "cache_read_input_token_cost": 1.5e-06, "deprecation_date": "2026-01-05", "input_cost_per_token": 1.5e-05, @@ -49007,14 +49007,14 @@ "supports_tool_choice": true }, "bedrock_mantle/openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "litellm_provider": "bedrock_mantle", "max_input_tokens": 1000000, "max_output_tokens": 128000, @@ -49070,6 +49070,34 @@ "supports_tool_choice": true, "supports_vision": true }, + "bedrock_mantle/openai.gpt-5.6-cyber": { + "input_cost_per_token": 1.375e-05, + "cache_creation_input_token_cost": 1.71875e-05, + "cache_read_input_token_cost": 1.375e-06, + "output_cost_per_token": 8.25e-05, + "litellm_provider": "bedrock_mantle", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "use_openai_responses_path": true, + "supported_endpoints": [ + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "bedrock_mantle/openai.gpt-5.6-luna": { "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, @@ -49103,14 +49131,14 @@ "supports_vision": true }, "us.openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "litellm_provider": "bedrock_converse", "max_input_tokens": 1000000, "max_output_tokens": 128000, @@ -49128,14 +49156,14 @@ "supports_vision": true }, "global.openai.gpt-5.6-sol": { - "input_cost_per_token": 5e-06, - "input_cost_per_token_above_272k_tokens": 1e-05, - "cache_creation_input_token_cost": 6.25e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, - "cache_read_input_token_cost": 5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1e-06, - "output_cost_per_token": 3e-05, - "output_cost_per_token_above_272k_tokens": 4.5e-05, + "input_cost_per_token": 4e-06, + "input_cost_per_token_above_272k_tokens": 8e-06, + "cache_creation_input_token_cost": 5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1e-05, + "cache_read_input_token_cost": 4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8e-07, + "output_cost_per_token": 2e-05, + "output_cost_per_token_above_272k_tokens": 3e-05, "litellm_provider": "bedrock_converse", "max_input_tokens": 1000000, "max_output_tokens": 128000, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 9f953e11df1..73e16715cf0 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -12282,7 +12282,7 @@ }, "claude-3-haiku-20240307": { "cache_creation_input_token_cost": 3e-07, - "cache_creation_input_token_cost_above_1hr": 6e-06, + "cache_creation_input_token_cost_above_1hr": 5e-07, "cache_read_input_token_cost": 3e-08, "deprecation_date": "2026-04-20", "input_cost_per_token": 2.5e-07, @@ -12301,7 +12301,7 @@ }, "claude-3-opus-20240229": { "cache_creation_input_token_cost": 1.875e-05, - "cache_creation_input_token_cost_above_1hr": 6e-06, + "cache_creation_input_token_cost_above_1hr": 3e-05, "cache_read_input_token_cost": 1.5e-06, "deprecation_date": "2026-01-05", "input_cost_per_token": 1.5e-05, @@ -49007,14 +49007,14 @@ "supports_tool_choice": true }, "bedrock_mantle/openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "litellm_provider": "bedrock_mantle", "max_input_tokens": 1000000, "max_output_tokens": 128000, @@ -49070,6 +49070,34 @@ "supports_tool_choice": true, "supports_vision": true }, + "bedrock_mantle/openai.gpt-5.6-cyber": { + "input_cost_per_token": 1.375e-05, + "cache_creation_input_token_cost": 1.71875e-05, + "cache_read_input_token_cost": 1.375e-06, + "output_cost_per_token": 8.25e-05, + "litellm_provider": "bedrock_mantle", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "use_openai_responses_path": true, + "supported_endpoints": [ + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "bedrock_mantle/openai.gpt-5.6-luna": { "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, @@ -49103,14 +49131,14 @@ "supports_vision": true }, "us.openai.gpt-5.6-sol": { - "input_cost_per_token": 5.5e-06, - "input_cost_per_token_above_272k_tokens": 1.1e-05, - "cache_creation_input_token_cost": 6.875e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, - "cache_read_input_token_cost": 5.5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "output_cost_per_token": 3.3e-05, - "output_cost_per_token_above_272k_tokens": 4.95e-05, + "input_cost_per_token": 4.4e-06, + "input_cost_per_token_above_272k_tokens": 8.8e-06, + "cache_creation_input_token_cost": 5.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.1e-05, + "cache_read_input_token_cost": 4.4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8.8e-07, + "output_cost_per_token": 2.2e-05, + "output_cost_per_token_above_272k_tokens": 3.3e-05, "litellm_provider": "bedrock_converse", "max_input_tokens": 1000000, "max_output_tokens": 128000, @@ -49128,14 +49156,14 @@ "supports_vision": true }, "global.openai.gpt-5.6-sol": { - "input_cost_per_token": 5e-06, - "input_cost_per_token_above_272k_tokens": 1e-05, - "cache_creation_input_token_cost": 6.25e-06, - "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, - "cache_read_input_token_cost": 5e-07, - "cache_read_input_token_cost_above_272k_tokens": 1e-06, - "output_cost_per_token": 3e-05, - "output_cost_per_token_above_272k_tokens": 4.5e-05, + "input_cost_per_token": 4e-06, + "input_cost_per_token_above_272k_tokens": 8e-06, + "cache_creation_input_token_cost": 5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1e-05, + "cache_read_input_token_cost": 4e-07, + "cache_read_input_token_cost_above_272k_tokens": 8e-07, + "output_cost_per_token": 2e-05, + "output_cost_per_token_above_272k_tokens": 3e-05, "litellm_provider": "bedrock_converse", "max_input_tokens": 1000000, "max_output_tokens": 128000, diff --git a/tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py b/tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py index dbd31c7e81b..694ed109025 100644 --- a/tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py +++ b/tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py @@ -59,17 +59,17 @@ class GptProfile(NamedTuple): GPT_5_6_PROFILES = [ GptProfile( model_id="us.openai.gpt-5.6-sol", - input_cost=5.5e-06, input_cost_above_272k=1.1e-05, - cache_write=6.875e-06, cache_write_above_272k=1.375e-05, - cache_read=5.5e-07, cache_read_above_272k=1.1e-06, - output_cost=3.3e-05, output_cost_above_272k=4.95e-05, + input_cost=4.4e-06, input_cost_above_272k=8.8e-06, + cache_write=5.5e-06, cache_write_above_272k=1.1e-05, + cache_read=4.4e-07, cache_read_above_272k=8.8e-07, + output_cost=2.2e-05, output_cost_above_272k=3.3e-05, ), GptProfile( model_id="global.openai.gpt-5.6-sol", - input_cost=5e-06, input_cost_above_272k=1e-05, - cache_write=6.25e-06, cache_write_above_272k=1.25e-05, - cache_read=5e-07, cache_read_above_272k=1e-06, - output_cost=3e-05, output_cost_above_272k=4.5e-05, + input_cost=4e-06, input_cost_above_272k=8e-06, + cache_write=5e-06, cache_write_above_272k=1e-05, + cache_read=4e-07, cache_read_above_272k=8e-07, + output_cost=2e-05, output_cost_above_272k=3e-05, ), GptProfile( model_id="us.openai.gpt-5.6-terra", @@ -221,7 +221,7 @@ def test_bedrock_gpt_5_6_above_272k_tier_applies_to_cost(local_model_cost_map): custom_llm_provider="bedrock", ) - assert cost == pytest.approx((300000 * 1.1e-05) + (1000 * 4.95e-05), rel=1e-9) + assert cost == pytest.approx((300000 * 8.8e-06) + (1000 * 3.3e-05), rel=1e-9) def test_bedrock_gpt_5_6_bills_cache_read_tokens(local_model_cost_map): @@ -241,10 +241,10 @@ def test_bedrock_gpt_5_6_bills_cache_read_tokens(local_model_cost_map): custom_llm_provider="bedrock", ) - expected = (2 * 5.5e-06) + (15609 * 5.5e-07) + (5 * 3.3e-05) + expected = (2 * 4.4e-06) + (15609 * 4.4e-07) + (5 * 2.2e-05) assert cost == pytest.approx(expected, rel=1e-9) # Without cache_read_input_token_cost the cached prefix bills at zero. - assert cost > (15611 * 5.5e-06) * 0.1 + assert cost > (15611 * 4.4e-06) * 0.1 def test_bedrock_gpt_5_6_bills_cache_write_tokens(local_model_cost_map): @@ -263,7 +263,7 @@ def test_bedrock_gpt_5_6_bills_cache_write_tokens(local_model_cost_map): custom_llm_provider="bedrock", ) - expected = (2 * 5.5e-06) + (15609 * 6.875e-06) + (5 * 3.3e-05) + expected = (2 * 4.4e-06) + (15609 * 5.5e-06) + (5 * 2.2e-05) assert cost == pytest.approx(expected, rel=1e-9) diff --git a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py index 9e05d48a18f..94c2d6ff6b3 100644 --- a/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py +++ b/tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py @@ -1506,10 +1506,19 @@ class TestBedrockMantleResponsesPricing: assert info["cache_read_input_token_cost"] == pytest.approx(2.75e-07) assert info["max_input_tokens"] == 272000 + def test_gpt_5_6_cyber_pricing_and_mode(self, local_cost_map): + info = litellm.get_model_info("bedrock_mantle/openai.gpt-5.6-cyber") + assert info["mode"] == "responses" + assert info["input_cost_per_token"] == pytest.approx(1.375e-05) + assert info["cache_creation_input_token_cost"] == pytest.approx(1.71875e-05) + assert info["cache_read_input_token_cost"] == pytest.approx(1.375e-06) + assert info["output_cost_per_token"] == pytest.approx(8.25e-05) + assert info["max_input_tokens"] == 272000 + @pytest.mark.parametrize( "model, input_cost, cache_creation_cost, cache_read_cost, output_cost", [ - ("openai.gpt-5.6-sol", 5.5e-06, 6.875e-06, 5.5e-07, 3.3e-05), + ("openai.gpt-5.6-sol", 4.4e-06, 5.5e-06, 4.4e-07, 2.2e-05), ("openai.gpt-5.6-terra", 2.2e-06, 2.75e-06, 2.2e-07, 1.32e-05), ("openai.gpt-5.6-luna", 2.2e-07, 2.75e-07, 2.2e-08, 1.32e-06), ], @@ -1532,7 +1541,7 @@ class TestBedrockMantleResponsesPricing: @pytest.mark.parametrize( "model, input_cost, output_cost", [ - ("openai.gpt-5.6-sol", 5.5e-06, 3.3e-05), + ("openai.gpt-5.6-sol", 4.4e-06, 2.2e-05), ("openai.gpt-5.6-terra", 2.2e-06, 1.32e-05), ("openai.gpt-5.6-luna", 2.2e-07, 1.32e-06), ], From 310591f63c283633199cbc0d3249f3d911d0220a Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 25 Aug 2026 14:05:07 +0000 Subject: [PATCH 08/44] test(model_prices): pin claude 3 1h cache write rates to 2x base input Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...test_anthropic_sonnet_1hr_cache_pricing.py | 53 +++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py b/tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py index f534b431508..11fcdf31dfc 100644 --- a/tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py +++ b/tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py @@ -87,3 +87,56 @@ def test_anthropic_sonnet_1hr_cache_write_pricing( ), f"{model_key}: long-context 1hr/5min ratio is {ratio_lc}, expected 1.6" else: assert "cache_creation_input_token_cost_above_1hr_above_200k_tokens" not in info + + +CLAUDE_3_EXPECTED = [ + ("claude-3-haiku-20240307", 5e-07), + ("claude-3-opus-20240229", 3e-05), +] + + +@pytest.mark.parametrize("model_key, expected_1hr", CLAUDE_3_EXPECTED) +def test_claude_3_1hr_cache_write_pricing(model_data, model_key, expected_1hr): + """Haiku 3 and Opus 3 both carried Sonnet's 6e-06 1hr rate, overbilling Haiku 3 + 1-hour cache writes 12x and underbilling Opus 3 5x.""" + info = model_data[model_key] + + assert info["cache_creation_input_token_cost_above_1hr"] == expected_1hr + + +@pytest.mark.parametrize("model_key, expected_1hr", CLAUDE_3_EXPECTED) +def test_backup_matches_main_for_claude_3_1hr_cache_write(model_key, expected_1hr): + json_path = os.path.join( + os.path.dirname(__file__), + "../../litellm/model_prices_and_context_window_backup.json", + ) + with open(json_path) as f: + backup = json.load(f) + + assert ( + backup[model_key]["cache_creation_input_token_cost_above_1hr"] == expected_1hr + ) + + +def test_first_party_anthropic_1hr_cache_writes_are_2x_base_input(model_data): + """Anthropic charges 1-hour cache writes at 2x base input for every first-party + model, so any entry that drifts off that multiple is a copy-paste error.""" + offenders = tuple( + ( + model_key, + info["input_cost_per_token"], + info["cache_creation_input_token_cost_above_1hr"], + ) + for model_key, info in model_data.items() + if isinstance(info, dict) + and info.get("litellm_provider") == "anthropic" + and info.get("input_cost_per_token") + and info.get("cache_creation_input_token_cost_above_1hr") + and abs( + info["cache_creation_input_token_cost_above_1hr"] + - 2 * info["input_cost_per_token"] + ) + > 1e-12 + ) + + assert offenders == (), f"1hr cache write is not 2x base input for: {offenders}" From fb15851f535dbb2bf68b85a0b73e3b4bf1065319 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 25 Aug 2026 19:16:14 +0000 Subject: [PATCH 09/44] fix(model_prices): verified Novita, DeepInfra, W&B, Gemini cache-read and Fireworks registry fixes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 2340 ++++++++++++++++- model_prices_and_context_window.json | 2340 ++++++++++++++++- 2 files changed, 4406 insertions(+), 274 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 87b4b08ea62..e133f31cdac 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -16023,12 +16023,13 @@ "max_tokens": 4096, "max_input_tokens": 4096, "max_output_tokens": 4096, - "input_cost_per_token": 8e-08, - "output_cost_per_token": 9e-08, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/NousResearch/Hermes-3-Llama-3.1-405B": { "max_tokens": 131072, @@ -16045,11 +16046,12 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 3e-07, - "output_cost_per_token": 3e-07, + "input_cost_per_token": 7e-07, + "output_cost_per_token": 7e-07, "litellm_provider": "deepinfra", "mode": "chat", - "supports_tool_choice": false + "supports_tool_choice": false, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/QwQ-32B": { "max_tokens": 131072, @@ -16066,12 +16068,13 @@ "max_tokens": 32768, "max_input_tokens": 32768, "max_output_tokens": 32768, - "input_cost_per_token": 1.2e-07, - "output_cost_per_token": 3.9e-07, + "input_cost_per_token": 3.6e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen2.5-7B-Instruct": { "max_tokens": 32768, @@ -16099,12 +16102,13 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 6e-08, + "input_cost_per_token": 1.2e-07, "output_cost_per_token": 2.4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-235B-A22B": { "max_tokens": 40960, @@ -16122,11 +16126,12 @@ "max_input_tokens": 262144, "max_output_tokens": 262144, "input_cost_per_token": 9e-08, - "output_cost_per_token": 6e-07, + "output_cost_per_token": 5.5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-235B-A22B-Thinking-2507": { "max_tokens": 262144, @@ -16143,23 +16148,25 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 8e-08, - "output_cost_per_token": 2.9e-07, + "input_cost_per_token": 1.2e-07, + "output_cost_per_token": 5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-32B": { "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 1e-07, + "input_cost_per_token": 8e-08, "output_cost_per_token": 2.8e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct": { "max_tokens": 262144, @@ -16176,23 +16183,27 @@ "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 2.9e-07, - "output_cost_per_token": 1.2e-06, + "input_cost_per_token": 3e-07, + "output_cost_per_token": 1e-06, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Next-80B-A3B-Instruct": { "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 1.4e-07, - "output_cost_per_token": 1.4e-06, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 1.1e-06, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Next-80B-A3B-Thinking": { "max_tokens": 262144, @@ -16219,11 +16230,12 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 6.5e-07, - "output_cost_per_token": 7.5e-07, + "input_cost_per_token": 8.5e-07, + "output_cost_per_token": 8.5e-07, "litellm_provider": "deepinfra", "mode": "chat", - "supports_tool_choice": false + "supports_tool_choice": false, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Sao10K/L3.3-70B-Euryale-v2.3": { "max_tokens": 131072, @@ -16349,36 +16361,41 @@ "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 3.8e-07, + "input_cost_per_token": 3.2e-07, "output_cost_per_token": 8.9e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3-0324": { "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 2.5e-07, - "output_cost_per_token": 8.8e-07, + "input_cost_per_token": 2.4e-07, + "output_cost_per_token": 9e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "cache_read_input_token_cost": 1.35e-07, + "supports_prompt_caching": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3.1": { "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 2.7e-07, - "output_cost_per_token": 1e-06, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 9.5e-07, "cache_read_input_token_cost": 2.16e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, "supports_reasoning": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3.1-Terminus": { "max_tokens": 163840, @@ -16432,33 +16449,36 @@ "max_input_tokens": 131072, "max_output_tokens": 131072, "input_cost_per_token": 5e-08, - "output_cost_per_token": 1e-07, + "output_cost_per_token": 1.5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/google/gemma-3-27b-it": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 9e-08, + "input_cost_per_token": 8e-08, "output_cost_per_token": 1.6e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/google/gemma-3-4b-it": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 4e-08, - "output_cost_per_token": 8e-08, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-3.2-11B-Vision-Instruct": { "max_tokens": 131072, @@ -16496,34 +16516,37 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 1.3e-07, - "output_cost_per_token": 3.9e-07, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3.2e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_function_calling": true, - "supports_tool_choice": true + "supports_tool_choice": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8": { "max_tokens": 1048576, "max_input_tokens": 1048576, "max_output_tokens": 1048576, - "input_cost_per_token": 1.5e-07, - "output_cost_per_token": 6e-07, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 8e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-4-Scout-17B-16E-Instruct": { "max_tokens": 327680, "max_input_tokens": 327680, "max_output_tokens": 327680, - "input_cost_per_token": 8e-08, + "input_cost_per_token": 1e-07, "output_cost_per_token": 3e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-Guard-3-8B": { "max_tokens": 131072, @@ -16571,12 +16594,13 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 1e-07, - "output_cost_per_token": 2.8e-07, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Meta-Llama-3.1-8B-Instruct": { "max_tokens": 131072, @@ -16594,11 +16618,12 @@ "max_input_tokens": 131072, "max_output_tokens": 131072, "input_cost_per_token": 2e-08, - "output_cost_per_token": 3e-08, + "output_cost_per_token": 4e-08, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/microsoft/WizardLM-2-8x22B": { "max_tokens": 65536, @@ -16625,12 +16650,13 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 2e-08, - "output_cost_per_token": 4e-08, + "input_cost_per_token": 1.9e-08, + "output_cost_per_token": 3e-08, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/mistralai/Mistral-Small-24B-Instruct-2501": { "max_tokens": 32768, @@ -16712,14 +16738,16 @@ }, "deepinfra/nvidia/NVIDIA-Nemotron-3.5-Lightning": { "max_input_tokens": 262144, - "input_cost_per_token": 5e-08, + "input_cost_per_token": 8e-08, "output_cost_per_token": 2e-07, "litellm_provider": "deepinfra", "mode": "chat", - "source": "https://deepinfra.com/nvidia/NVIDIA-Nemotron-3.5-Lightning", + "source": "https://deepinfra.com/pricing", "supports_tool_choice": true, "supports_function_calling": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true }, "deepinfra/nvidia/NVIDIA-Nemotron-Nano-9B-v2": { "max_tokens": 131072, @@ -16736,23 +16764,25 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 5e-08, - "output_cost_per_token": 4.5e-07, + "input_cost_per_token": 3.7e-08, + "output_cost_per_token": 1.7e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/openai/gpt-oss-20b": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 4e-08, - "output_cost_per_token": 1.5e-07, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/zai-org/GLM-4.5": { "max_tokens": 131072, @@ -20237,7 +20267,7 @@ "supports_image_size": false }, "gemini-2.5-flash-preview-09-2025": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "vertex_ai-language-models", @@ -20247,7 +20277,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -20373,7 +20403,7 @@ }, "gemini-2.5-flash-lite-preview-06-17": { "deprecation_date": "2025-11-18", - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 5e-07, "input_cost_per_token": 1e-07, "litellm_provider": "vertex_ai-language-models", @@ -20383,7 +20413,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -21953,7 +21983,7 @@ "supports_image_size": false }, "gemini/gemini-2.5-flash-preview-09-2025": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "deprecation_date": "2026-02-17", "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, @@ -21965,7 +21995,7 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22001,7 +22031,7 @@ "supports_image_size": false }, "gemini/gemini-flash-latest": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "gemini", @@ -22012,7 +22042,7 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22047,7 +22077,7 @@ } }, "gemini/gemini-flash-lite-latest": { - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 3e-07, "input_cost_per_token": 1e-07, "litellm_provider": "gemini", @@ -22058,7 +22088,7 @@ "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22094,7 +22124,7 @@ }, "gemini/gemini-2.5-flash-lite-preview-06-17": { "deprecation_date": "2025-11-18", - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 5e-07, "input_cost_per_token": 1e-07, "litellm_provider": "gemini", @@ -22105,7 +22135,7 @@ "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, "rpm": 15, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-lite", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -42511,19 +42541,21 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 0.015, - "output_cost_per_token": 0.06, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.7e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/openai/gpt-oss-20b": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 0.005, - "output_cost_per_token": 0.02, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.3e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/zai-org/GLM-4.5": { "max_tokens": 131072, @@ -42547,10 +42579,11 @@ "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 0.1, - "output_cost_per_token": 0.15, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 1.5e-06, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/Qwen/Qwen3-235B-A22B-Thinking-2507": { "max_tokens": 262144, @@ -42602,19 +42635,21 @@ "max_tokens": 128000, "max_input_tokens": 128000, "max_output_tokens": 128000, - "input_cost_per_token": 0.022, - "output_cost_per_token": 0.022, + "input_cost_per_token": 2.2e-07, + "output_cost_per_token": 2.2e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/deepseek-ai/DeepSeek-V3.1": { "max_tokens": 128000, - "max_input_tokens": 128000, + "max_input_tokens": 161000, "max_output_tokens": 128000, - "input_cost_per_token": 0.055, - "output_cost_per_token": 0.165, + "input_cost_per_token": 5.5e-07, + "output_cost_per_token": 1.65e-06, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/deepseek-ai/DeepSeek-R1-0528": { "max_tokens": 161000, @@ -42638,10 +42673,11 @@ "max_tokens": 128000, "max_input_tokens": 128000, "max_output_tokens": 128000, - "input_cost_per_token": 0.071, - "output_cost_per_token": 0.071, + "input_cost_per_token": 7.1e-07, + "output_cost_per_token": 7.1e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/meta-llama/Llama-4-Scout-17B-16E-Instruct": { "max_tokens": 64000, @@ -46608,8 +46644,8 @@ "novita/xiaomimimo/mimo-v2-flash": { "litellm_provider": "novita", "mode": "chat", - "input_cost_per_token": 1e-07, - "output_cost_per_token": 3e-07, + "input_cost_per_token": 1.1e-07, + "output_cost_per_token": 3.3e-07, "max_input_tokens": 262144, "max_output_tokens": 32000, "max_tokens": 32000, @@ -46618,8 +46654,8 @@ "supports_tool_choice": true, "supports_system_messages": true, "supports_response_schema": true, - "cache_read_input_token_cost": 2e-08, - "input_cost_per_token_cache_hit": 2e-08, + "cache_read_input_token_cost": 2.4e-08, + "input_cost_per_token_cache_hit": 2.4e-08, "supports_reasoning": true }, "novita/zai-org/autoglm-phone-9b-multilingual": { @@ -46639,14 +46675,16 @@ "input_cost_per_token": 6e-07, "output_cost_per_token": 2.5e-06, "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, "supports_response_schema": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true }, "novita/minimax/minimax-m2": { "litellm_provider": "novita", @@ -46662,7 +46700,8 @@ "supports_system_messages": true, "cache_read_input_token_cost": 3e-08, "input_cost_per_token_cache_hit": 3e-08, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/paddlepaddle/paddleocr-vl": { "litellm_provider": "novita", @@ -46700,7 +46739,9 @@ "max_tokens": 32768, "supports_vision": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/zai-org/glm-4.6v": { "litellm_provider": "novita", @@ -46765,7 +46806,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_response_schema": true + "supports_response_schema": true, + "supports_reasoning": true }, "novita/qwen/qwen3-next-80b-a3b-thinking": { "litellm_provider": "novita", @@ -46877,8 +46919,8 @@ "input_cost_per_token": 6e-07, "output_cost_per_token": 2.5e-06, "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46888,8 +46930,8 @@ "novita/qwen/qwen3-coder-480b-a35b-instruct": { "litellm_provider": "novita", "mode": "chat", - "input_cost_per_token": 3e-07, - "output_cost_per_token": 1.3e-06, + "input_cost_per_token": 3.8e-07, + "output_cost_per_token": 1.55e-06, "max_input_tokens": 262144, "max_output_tokens": 65536, "max_tokens": 65536, @@ -46935,8 +46977,8 @@ "input_cost_per_token": 5.7e-07, "output_cost_per_token": 2.3e-06, "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46949,8 +46991,8 @@ "input_cost_per_token": 2.7e-07, "output_cost_per_token": 1.12e-06, "max_input_tokens": 163840, - "max_output_tokens": 163840, - "max_tokens": 163840, + "max_output_tokens": 65536, + "max_tokens": 65536, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46997,7 +47039,8 @@ "max_input_tokens": 16384, "max_output_tokens": 16384, "max_tokens": 16384, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/google/gemma-3-12b-it": { "litellm_provider": "novita", @@ -47076,13 +47119,14 @@ "mode": "chat", "input_cost_per_token": 1.35e-07, "output_cost_per_token": 4e-07, - "max_input_tokens": 131072, - "max_output_tokens": 120000, - "max_tokens": 120000, + "max_input_tokens": 12288, + "max_output_tokens": 12288, + "max_tokens": 12288, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/qwen/qwen-2.5-72b-instruct": { "litellm_provider": "novita", @@ -47122,7 +47166,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/deepseek/deepseek-r1-0528": { "litellm_provider": "novita", @@ -47162,7 +47207,8 @@ "max_input_tokens": 8192, "max_output_tokens": 8192, "max_tokens": 8192, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/microsoft/wizardlm-2-8x22b": { "litellm_provider": "novita", @@ -47172,7 +47218,8 @@ "max_input_tokens": 65535, "max_output_tokens": 8000, "max_tokens": 8000, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/deepseek/deepseek-r1-0528-qwen3-8b": { "litellm_provider": "novita", @@ -47219,7 +47266,8 @@ "max_output_tokens": 20000, "max_tokens": 20000, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8": { "litellm_provider": "novita", @@ -47230,7 +47278,8 @@ "max_output_tokens": 8192, "max_tokens": 8192, "supports_vision": true, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/meta-llama/llama-4-scout-17b-16e-instruct": { "litellm_provider": "novita", @@ -47366,7 +47415,9 @@ "max_output_tokens": 20000, "max_tokens": 20000, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/google/gemma-3-27b-it": { "litellm_provider": "novita", @@ -47404,7 +47455,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/Sao10K/L3-8B-Stheno-v3.2": { "litellm_provider": "novita", @@ -47472,7 +47524,9 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 2.5e-08, + "supports_prompt_caching": true }, "novita/qwen/qwen3-vl-30b-a3b-instruct": { "litellm_provider": "novita", @@ -47593,10 +47647,12 @@ "input_cost_per_token": 3e-08, "output_cost_per_token": 3e-08, "max_input_tokens": 128000, - "max_output_tokens": 20000, - "max_tokens": 20000, + "max_output_tokens": 8192, + "max_tokens": 8192, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/qwen/qwen2.5-7b-instruct": { "litellm_provider": "novita", @@ -47604,8 +47660,8 @@ "input_cost_per_token": 7e-08, "output_cost_per_token": 7e-08, "max_input_tokens": 32000, - "max_output_tokens": 32000, - "max_tokens": 32000, + "max_output_tokens": 8192, + "max_tokens": 8192, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -50838,14 +50894,14 @@ "supports_embedding_image_input": true }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731": { - "cache_read_input_token_cost": 2.8e-08, - "input_cost_per_token": 1.4e-07, + "cache_read_input_token_cost": 7e-09, + "input_cost_per_token": 2.2e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", - "output_cost_per_token": 2.8e-07, + "output_cost_per_token": 6.6e-07, "source": "https://docs.fireworks.ai/serverless/pricing", "supports_function_calling": true, "supports_reasoning": true, @@ -51152,5 +51208,2015 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true + }, + "novita/zai-org/glm-5.3": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v4-pro-0813": { + "cache_read_input_token_cost": 1.3200000000000002e-07, + "input_cost_per_token": 1.32e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 3.96e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/moonshotai/kimi-k3": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 3e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 1048576, + "max_tokens": 1048576, + "mode": "chat", + "output_cost_per_token": 1.5e-05, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/tencent/hy3": { + "cache_read_input_token_cost": 3.5e-08, + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 5.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5.2": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/moonshotai/kimi-k2.7-code": { + "cache_read_input_token_cost": 1.9e-07, + "input_cost_per_token": 9.499999999999999e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash-vision-exp": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 4.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 1.32e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash-0731": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 4.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 1.32e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/mindai/macaron-v1-venti": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 1.5e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.5e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m3": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 2.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v4-pro": { + "cache_read_input_token_cost": 1.35e-07, + "input_cost_per_token": 1.6000000000000001e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/inclusionai/ling-3.0-flash-fast": { + "cache_read_input_token_cost": 1.2e-08, + "input_cost_per_token": 6e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.8-max": { + "cache_read_input_token_cost": 2.5e-07, + "input_cost_per_token": 2e-06, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 6e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/inclusionai/ling-3.0-flash": { + "cache_read_input_token_cost": 1.2e-08, + "input_cost_per_token": 6e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/mindai/macaron-v1-tall": { + "cache_read_input_token_cost": 8e-08, + "input_cost_per_token": 4.5000000000000003e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 2.6e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/stepfun/step-3.7-flash": { + "cache_read_input_token_cost": 4e-08, + "input_cost_per_token": 2.0000000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 256000, + "max_tokens": 256000, + "mode": "chat", + "output_cost_per_token": 1.15e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/nvidia/nemotron-3-nano-30b-a3b": { + "input_cost_per_token": 5.0000000000000004e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 2.0000000000000002e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/baidu/cobuddy": { + "cache_read_input_token_cost": 7e-08, + "input_cost_per_token": 2.8e-07, + "litellm_provider": "novita", + "max_input_tokens": 131072, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.13e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/xiaomimimo/mimo-v2.5": { + "cache_read_input_token_cost": 3.4e-09, + "input_cost_per_token": 1.6800000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.3600000000000004e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.7-max": { + "cache_read_input_token_cost": 2.5e-07, + "input_cost_per_token": 1.25e-06, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.75e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/xiaomimimo/mimo-v2.5-pro": { + "cache_read_input_token_cost": 4.3e-09, + "input_cost_per_token": 5.22e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.044e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.6-27b": { + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.6000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/moonshotai/kimi-k2.6": { + "cache_read_input_token_cost": 1.6e-07, + "input_cost_per_token": 8.000000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 3.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-5.1": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.38e-06, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.7-highspeed": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5v-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/google/gemma-4-26b-a4b-it": { + "input_cost_per_token": 1.3e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/google/gemma-4-31b-it": { + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-5-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "novita", + "max_input_tokens": 202800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.7": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.5-highspeed": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131100, + "max_tokens": 131100, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.5-27b": { + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-122b-a10b": { + "input_cost_per_token": 4.0000000000000003e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-35b-a3b": { + "input_cost_per_token": 2.5e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-397b-a17b": { + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.6000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/minimax/minimax-m2.5": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131100, + "max_tokens": 131100, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5": { + "cache_read_input_token_cost": 2.0000000000000002e-07, + "input_cost_per_token": 1e-06, + "litellm_provider": "novita", + "max_input_tokens": 202800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3-coder-next": { + "input_cost_per_token": 2.0000000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.5e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-ocr-2": { + "input_cost_per_token": 3e-08, + "litellm_provider": "novita", + "max_input_tokens": 8192, + "max_output_tokens": 8192, + "max_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 3e-08, + "source": "https://novita.ai/pricing", + "supports_vision": true + }, + "novita/moonshotai/kimi-k2.5": { + "cache_read_input_token_cost": 1.0000000000000001e-07, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 3e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-4.7-h": { + "cache_read_input_token_cost": 1.1e-07, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 2.2e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-4.7-flash": { + "cache_read_input_token_cost": 1e-08, + "input_cost_per_token": 7e-08, + "litellm_provider": "novita", + "max_input_tokens": 200000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.6-35b-a3b": { + "input_cost_per_token": 2.48e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.4850000000000002e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek_v3": { + "input_cost_per_token": 8.900000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 16000, + "max_tokens": 16000, + "mode": "chat", + "output_cost_per_token": 8.900000000000001e-07, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-r1": { + "input_cost_per_token": 4e-06, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 16000, + "max_tokens": 16000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v3/community": { + "input_cost_per_token": 8.900000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 8000, + "max_tokens": 8000, + "mode": "chat", + "output_cost_per_token": 8.900000000000001e-07, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-r1/community": { + "input_cost_per_token": 4e-06, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 8000, + "max_tokens": 8000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/thudm/glm-4-32b-0414": { + "input_cost_per_token": 5.5e-07, + "litellm_provider": "novita", + "max_input_tokens": 32000, + "max_output_tokens": 32000, + "max_tokens": 32000, + "mode": "chat", + "output_cost_per_token": 1.66e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/meta-llama/llama-3.2-1b-instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "novita", + "max_input_tokens": 131000, + "max_output_tokens": 32000, + "max_tokens": 32000, + "mode": "chat", + "output_cost_per_token": 2e-08, + "source": "https://api.novita.ai/v3/openai/models", + "supports_response_schema": true, + "supports_vision": false + }, + "wandb/deepseek-ai/DeepSeek-V4-Flash": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 2.8e-07, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/deepseek-ai/DeepSeek-V4-Flash-0731": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.3e-07, + "output_cost_per_token": 2.8e-07, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/deepseek-ai/DeepSeek-V4-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.15e-06, + "output_cost_per_token": 2.55e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/google/gemma-4-31B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3.4e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/ibm-granite/granite-4.1-8b": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/JetBrains/Mellum2-12B-A2.5B-Instruct": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/meta-llama/Llama-3.1-70B-Instruct": { + "max_tokens": 128000, + "max_input_tokens": 128000, + "input_cost_per_token": 8e-07, + "output_cost_per_token": 8e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/MiniMaxAI/MiniMax-M3": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.3e-07, + "output_cost_per_token": 9.6e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/moonshotai/Kimi-K2.7-Code": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.1e-07, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/moonshotai/Kimi-K2.6": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6.5e-07, + "output_cost_per_token": 3.41e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 2.5e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 2.75e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/OpenPipe/Qwen3-14B-Instruct": { + "max_tokens": 32768, + "max_input_tokens": 32768, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 2.2e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.8-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.6-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.25e-06, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.6-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6e-07, + "output_cost_per_token": 3.6e-06, + "cache_read_input_token_cost": 1.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.5-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.25e-06, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3-30B-A3B-Instruct-2507": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/zai-org/GLM-5.2": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.6e-07, + "output_cost_per_token": 2.42e-06, + "cache_read_input_token_cost": 1.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "deepinfra/openai/gpt-oss-120b-Turbo": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 1.5e-07, + "output_cost_per_token": 6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M2.7": { + "max_tokens": 196608, + "max_input_tokens": 196608, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it-Ultra": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2.7e-07, + "output_cost_per_token": 7.6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.5": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 2.25e-06, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.7-Flash": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 6e-08, + "output_cost_per_token": 4e-07, + "cache_read_input_token_cost": 1e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.6": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-4-8": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-sonnet-4-6": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 3e-06, + "output_cost_per_token": 1.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.5-flash": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 1.5e-06, + "output_cost_per_token": 9e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/XiaomiMiMo/MiMo-V2.5": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it-turbo": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 3.4e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/thinkingmachines/Inkling-Small": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 1.2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/meta-models/Muse-Glimmer-30B": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 3e-07, + "output_cost_per_token": 1.2e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-Max-Thinking": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-VL-235B-A22B-Instruct": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 8.8e-07, + "cache_read_input_token_cost": 1.1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-VL-30B-A3B-Instruct": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.5e-07, + "output_cost_per_token": 6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.6e-07, + "output_cost_per_token": 2.6e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.6-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 9.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/Nemotron-Content-Safety-3.5": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 2e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/thinkingmachines/Inkling": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 9.5e-07, + "output_cost_per_token": 4.05e-06, + "cache_read_input_token_cost": 1.6e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.6": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Pro-0813": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.3e-06, + "output_cost_per_token": 2.6e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.7-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 2.5e-06, + "output_cost_per_token": 7.5e-06, + "cache_read_input_token_cost": 5e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-mini": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 4e-07, + "cache_read_input_token_cost": 2e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-2.4T-A95B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M3": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 2.8e-07, + "output_cost_per_token": 1.1e-06, + "cache_read_input_token_cost": 5.6e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.1-flash-lite": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.5e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.7-flash": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 3.75e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/inclusionAI/Ling-3.0-flash": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 6e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.2e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/stepfun-ai/Step-3.7-Flash": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 1.15e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 1e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-1.8": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/tencent/Hy3": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 5.8e-07, + "cache_read_input_token_cost": 3.5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-code": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-pro": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 6e-07, + "output_cost_per_token": 2.08e-06, + "cache_read_input_token_cost": 1.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/Nemotron-3-Nano-30B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 2e-07, + "cache_read_input_token_cost": 2.5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.7-Code": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6.8e-07, + "output_cost_per_token": 3.4e-06, + "cache_read_input_token_cost": 1.36e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-sonnet-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 1e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-397B-A17B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 2.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Flash-0731": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 8e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.6e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-E4B-it": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V3.2": { + "max_tokens": 163840, + "max_input_tokens": 163840, + "input_cost_per_token": 2.6e-07, + "output_cost_per_token": 3.8e-07, + "cache_read_input_token_cost": 1.3e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.65e-06, + "output_cost_per_token": 4.951e-06, + "cache_read_input_token_cost": 2.06e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-fable-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 1e-05, + "output_cost_per_token": 5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 2.2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-122B-A10B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.9e-07, + "output_cost_per_token": 2.4e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5.1": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 1.05e-06, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 2.05e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.3e-06, + "output_cost_per_token": 2.6e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 8.5e-08, + "output_cost_per_token": 4e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5.2": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 2.4e-06, + "cache_read_input_token_cost": 1.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K3": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 2.85e-06, + "output_cost_per_token": 1.425e-05, + "cache_read_input_token_cost": 2.85e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-4-7": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.6-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 3.2e-07, + "output_cost_per_token": 3.2e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-26B-A4B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7e-08, + "output_cost_per_token": 3.4e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.1-pro": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 1.2e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/XiaomiMiMo/MiMo-V2.5-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-haiku-4-5": { + "max_tokens": 200000, + "max_input_tokens": 200000, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 5e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Flash": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/openai/gpt-oss-120b-Ultra": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 9.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-9B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 1.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M2.7-Turbo": { + "max_tokens": 196608, + "max_input_tokens": 196608, + "input_cost_per_token": 3.8e-07, + "output_cost_per_token": 1.7e-06, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.7": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 1.75e-06, + "cache_read_input_token_cost": 8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.3e-07, + "output_cost_per_token": 3.8e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" } } diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 87b4b08ea62..e133f31cdac 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -16023,12 +16023,13 @@ "max_tokens": 4096, "max_input_tokens": 4096, "max_output_tokens": 4096, - "input_cost_per_token": 8e-08, - "output_cost_per_token": 9e-08, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/NousResearch/Hermes-3-Llama-3.1-405B": { "max_tokens": 131072, @@ -16045,11 +16046,12 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 3e-07, - "output_cost_per_token": 3e-07, + "input_cost_per_token": 7e-07, + "output_cost_per_token": 7e-07, "litellm_provider": "deepinfra", "mode": "chat", - "supports_tool_choice": false + "supports_tool_choice": false, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/QwQ-32B": { "max_tokens": 131072, @@ -16066,12 +16068,13 @@ "max_tokens": 32768, "max_input_tokens": 32768, "max_output_tokens": 32768, - "input_cost_per_token": 1.2e-07, - "output_cost_per_token": 3.9e-07, + "input_cost_per_token": 3.6e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen2.5-7B-Instruct": { "max_tokens": 32768, @@ -16099,12 +16102,13 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 6e-08, + "input_cost_per_token": 1.2e-07, "output_cost_per_token": 2.4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-235B-A22B": { "max_tokens": 40960, @@ -16122,11 +16126,12 @@ "max_input_tokens": 262144, "max_output_tokens": 262144, "input_cost_per_token": 9e-08, - "output_cost_per_token": 6e-07, + "output_cost_per_token": 5.5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-235B-A22B-Thinking-2507": { "max_tokens": 262144, @@ -16143,23 +16148,25 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 8e-08, - "output_cost_per_token": 2.9e-07, + "input_cost_per_token": 1.2e-07, + "output_cost_per_token": 5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-32B": { "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 1e-07, + "input_cost_per_token": 8e-08, "output_cost_per_token": 2.8e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct": { "max_tokens": 262144, @@ -16176,23 +16183,27 @@ "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 2.9e-07, - "output_cost_per_token": 1.2e-06, + "input_cost_per_token": 3e-07, + "output_cost_per_token": 1e-06, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Next-80B-A3B-Instruct": { "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 1.4e-07, - "output_cost_per_token": 1.4e-06, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 1.1e-06, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Qwen/Qwen3-Next-80B-A3B-Thinking": { "max_tokens": 262144, @@ -16219,11 +16230,12 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 6.5e-07, - "output_cost_per_token": 7.5e-07, + "input_cost_per_token": 8.5e-07, + "output_cost_per_token": 8.5e-07, "litellm_provider": "deepinfra", "mode": "chat", - "supports_tool_choice": false + "supports_tool_choice": false, + "source": "https://deepinfra.com/pricing" }, "deepinfra/Sao10K/L3.3-70B-Euryale-v2.3": { "max_tokens": 131072, @@ -16349,36 +16361,41 @@ "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 3.8e-07, + "input_cost_per_token": 3.2e-07, "output_cost_per_token": 8.9e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3-0324": { "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 2.5e-07, - "output_cost_per_token": 8.8e-07, + "input_cost_per_token": 2.4e-07, + "output_cost_per_token": 9e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "cache_read_input_token_cost": 1.35e-07, + "supports_prompt_caching": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3.1": { "max_tokens": 163840, "max_input_tokens": 163840, "max_output_tokens": 163840, - "input_cost_per_token": 2.7e-07, - "output_cost_per_token": 1e-06, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 9.5e-07, "cache_read_input_token_cost": 2.16e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, "supports_reasoning": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/deepseek-ai/DeepSeek-V3.1-Terminus": { "max_tokens": 163840, @@ -16432,33 +16449,36 @@ "max_input_tokens": 131072, "max_output_tokens": 131072, "input_cost_per_token": 5e-08, - "output_cost_per_token": 1e-07, + "output_cost_per_token": 1.5e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/google/gemma-3-27b-it": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 9e-08, + "input_cost_per_token": 8e-08, "output_cost_per_token": 1.6e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/google/gemma-3-4b-it": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 4e-08, - "output_cost_per_token": 8e-08, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-3.2-11B-Vision-Instruct": { "max_tokens": 131072, @@ -16496,34 +16516,37 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 1.3e-07, - "output_cost_per_token": 3.9e-07, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3.2e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_function_calling": true, - "supports_tool_choice": true + "supports_tool_choice": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8": { "max_tokens": 1048576, "max_input_tokens": 1048576, "max_output_tokens": 1048576, - "input_cost_per_token": 1.5e-07, - "output_cost_per_token": 6e-07, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 8e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-4-Scout-17B-16E-Instruct": { "max_tokens": 327680, "max_input_tokens": 327680, "max_output_tokens": 327680, - "input_cost_per_token": 8e-08, + "input_cost_per_token": 1e-07, "output_cost_per_token": 3e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Llama-Guard-3-8B": { "max_tokens": 131072, @@ -16571,12 +16594,13 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 1e-07, - "output_cost_per_token": 2.8e-07, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/meta-llama/Meta-Llama-3.1-8B-Instruct": { "max_tokens": 131072, @@ -16594,11 +16618,12 @@ "max_input_tokens": 131072, "max_output_tokens": 131072, "input_cost_per_token": 2e-08, - "output_cost_per_token": 3e-08, + "output_cost_per_token": 4e-08, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/microsoft/WizardLM-2-8x22B": { "max_tokens": 65536, @@ -16625,12 +16650,13 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 2e-08, - "output_cost_per_token": 4e-08, + "input_cost_per_token": 1.9e-08, + "output_cost_per_token": 3e-08, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/mistralai/Mistral-Small-24B-Instruct-2501": { "max_tokens": 32768, @@ -16712,14 +16738,16 @@ }, "deepinfra/nvidia/NVIDIA-Nemotron-3.5-Lightning": { "max_input_tokens": 262144, - "input_cost_per_token": 5e-08, + "input_cost_per_token": 8e-08, "output_cost_per_token": 2e-07, "litellm_provider": "deepinfra", "mode": "chat", - "source": "https://deepinfra.com/nvidia/NVIDIA-Nemotron-3.5-Lightning", + "source": "https://deepinfra.com/pricing", "supports_tool_choice": true, "supports_function_calling": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true }, "deepinfra/nvidia/NVIDIA-Nemotron-Nano-9B-v2": { "max_tokens": 131072, @@ -16736,23 +16764,25 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 5e-08, - "output_cost_per_token": 4.5e-07, + "input_cost_per_token": 3.7e-08, + "output_cost_per_token": 1.7e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/openai/gpt-oss-20b": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 4e-08, - "output_cost_per_token": 1.5e-07, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.4e-07, "litellm_provider": "deepinfra", "mode": "chat", "supports_tool_choice": true, - "supports_function_calling": true + "supports_function_calling": true, + "source": "https://deepinfra.com/pricing" }, "deepinfra/zai-org/GLM-4.5": { "max_tokens": 131072, @@ -20237,7 +20267,7 @@ "supports_image_size": false }, "gemini-2.5-flash-preview-09-2025": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "vertex_ai-language-models", @@ -20247,7 +20277,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -20373,7 +20403,7 @@ }, "gemini-2.5-flash-lite-preview-06-17": { "deprecation_date": "2025-11-18", - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 5e-07, "input_cost_per_token": 1e-07, "litellm_provider": "vertex_ai-language-models", @@ -20383,7 +20413,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -21953,7 +21983,7 @@ "supports_image_size": false }, "gemini/gemini-2.5-flash-preview-09-2025": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "deprecation_date": "2026-02-17", "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, @@ -21965,7 +21995,7 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22001,7 +22031,7 @@ "supports_image_size": false }, "gemini/gemini-flash-latest": { - "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "gemini", @@ -22012,7 +22042,7 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22047,7 +22077,7 @@ } }, "gemini/gemini-flash-lite-latest": { - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 3e-07, "input_cost_per_token": 1e-07, "litellm_provider": "gemini", @@ -22058,7 +22088,7 @@ "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, "rpm": 15, - "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -22094,7 +22124,7 @@ }, "gemini/gemini-2.5-flash-lite-preview-06-17": { "deprecation_date": "2025-11-18", - "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 5e-07, "input_cost_per_token": 1e-07, "litellm_provider": "gemini", @@ -22105,7 +22135,7 @@ "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, "rpm": 15, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-lite", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -42511,19 +42541,21 @@ "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 0.015, - "output_cost_per_token": 0.06, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.7e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/openai/gpt-oss-20b": { "max_tokens": 131072, "max_input_tokens": 131072, "max_output_tokens": 131072, - "input_cost_per_token": 0.005, - "output_cost_per_token": 0.02, + "input_cost_per_token": 3e-08, + "output_cost_per_token": 1.3e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/zai-org/GLM-4.5": { "max_tokens": 131072, @@ -42547,10 +42579,11 @@ "max_tokens": 262144, "max_input_tokens": 262144, "max_output_tokens": 262144, - "input_cost_per_token": 0.1, - "output_cost_per_token": 0.15, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 1.5e-06, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/Qwen/Qwen3-235B-A22B-Thinking-2507": { "max_tokens": 262144, @@ -42602,19 +42635,21 @@ "max_tokens": 128000, "max_input_tokens": 128000, "max_output_tokens": 128000, - "input_cost_per_token": 0.022, - "output_cost_per_token": 0.022, + "input_cost_per_token": 2.2e-07, + "output_cost_per_token": 2.2e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/deepseek-ai/DeepSeek-V3.1": { "max_tokens": 128000, - "max_input_tokens": 128000, + "max_input_tokens": 161000, "max_output_tokens": 128000, - "input_cost_per_token": 0.055, - "output_cost_per_token": 0.165, + "input_cost_per_token": 5.5e-07, + "output_cost_per_token": 1.65e-06, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/deepseek-ai/DeepSeek-R1-0528": { "max_tokens": 161000, @@ -42638,10 +42673,11 @@ "max_tokens": 128000, "max_input_tokens": 128000, "max_output_tokens": 128000, - "input_cost_per_token": 0.071, - "output_cost_per_token": 0.071, + "input_cost_per_token": 7.1e-07, + "output_cost_per_token": 7.1e-07, "litellm_provider": "wandb", - "mode": "chat" + "mode": "chat", + "source": "https://wandb.ai/site/pricing/tokens/" }, "wandb/meta-llama/Llama-4-Scout-17B-16E-Instruct": { "max_tokens": 64000, @@ -46608,8 +46644,8 @@ "novita/xiaomimimo/mimo-v2-flash": { "litellm_provider": "novita", "mode": "chat", - "input_cost_per_token": 1e-07, - "output_cost_per_token": 3e-07, + "input_cost_per_token": 1.1e-07, + "output_cost_per_token": 3.3e-07, "max_input_tokens": 262144, "max_output_tokens": 32000, "max_tokens": 32000, @@ -46618,8 +46654,8 @@ "supports_tool_choice": true, "supports_system_messages": true, "supports_response_schema": true, - "cache_read_input_token_cost": 2e-08, - "input_cost_per_token_cache_hit": 2e-08, + "cache_read_input_token_cost": 2.4e-08, + "input_cost_per_token_cache_hit": 2.4e-08, "supports_reasoning": true }, "novita/zai-org/autoglm-phone-9b-multilingual": { @@ -46639,14 +46675,16 @@ "input_cost_per_token": 6e-07, "output_cost_per_token": 2.5e-06, "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, "supports_response_schema": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true }, "novita/minimax/minimax-m2": { "litellm_provider": "novita", @@ -46662,7 +46700,8 @@ "supports_system_messages": true, "cache_read_input_token_cost": 3e-08, "input_cost_per_token_cache_hit": 3e-08, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/paddlepaddle/paddleocr-vl": { "litellm_provider": "novita", @@ -46700,7 +46739,9 @@ "max_tokens": 32768, "supports_vision": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/zai-org/glm-4.6v": { "litellm_provider": "novita", @@ -46765,7 +46806,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_response_schema": true + "supports_response_schema": true, + "supports_reasoning": true }, "novita/qwen/qwen3-next-80b-a3b-thinking": { "litellm_provider": "novita", @@ -46877,8 +46919,8 @@ "input_cost_per_token": 6e-07, "output_cost_per_token": 2.5e-06, "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46888,8 +46930,8 @@ "novita/qwen/qwen3-coder-480b-a35b-instruct": { "litellm_provider": "novita", "mode": "chat", - "input_cost_per_token": 3e-07, - "output_cost_per_token": 1.3e-06, + "input_cost_per_token": 3.8e-07, + "output_cost_per_token": 1.55e-06, "max_input_tokens": 262144, "max_output_tokens": 65536, "max_tokens": 65536, @@ -46935,8 +46977,8 @@ "input_cost_per_token": 5.7e-07, "output_cost_per_token": 2.3e-06, "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, + "max_output_tokens": 100352, + "max_tokens": 100352, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46949,8 +46991,8 @@ "input_cost_per_token": 2.7e-07, "output_cost_per_token": 1.12e-06, "max_input_tokens": 163840, - "max_output_tokens": 163840, - "max_tokens": 163840, + "max_output_tokens": 65536, + "max_tokens": 65536, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -46997,7 +47039,8 @@ "max_input_tokens": 16384, "max_output_tokens": 16384, "max_tokens": 16384, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/google/gemma-3-12b-it": { "litellm_provider": "novita", @@ -47076,13 +47119,14 @@ "mode": "chat", "input_cost_per_token": 1.35e-07, "output_cost_per_token": 4e-07, - "max_input_tokens": 131072, - "max_output_tokens": 120000, - "max_tokens": 120000, + "max_input_tokens": 12288, + "max_output_tokens": 12288, + "max_tokens": 12288, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/qwen/qwen-2.5-72b-instruct": { "litellm_provider": "novita", @@ -47122,7 +47166,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/deepseek/deepseek-r1-0528": { "litellm_provider": "novita", @@ -47162,7 +47207,8 @@ "max_input_tokens": 8192, "max_output_tokens": 8192, "max_tokens": 8192, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/microsoft/wizardlm-2-8x22b": { "litellm_provider": "novita", @@ -47172,7 +47218,8 @@ "max_input_tokens": 65535, "max_output_tokens": 8000, "max_tokens": 8000, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/deepseek/deepseek-r1-0528-qwen3-8b": { "litellm_provider": "novita", @@ -47219,7 +47266,8 @@ "max_output_tokens": 20000, "max_tokens": 20000, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8": { "litellm_provider": "novita", @@ -47230,7 +47278,8 @@ "max_output_tokens": 8192, "max_tokens": 8192, "supports_vision": true, - "supports_system_messages": true + "supports_system_messages": true, + "supports_response_schema": true }, "novita/meta-llama/llama-4-scout-17b-16e-instruct": { "litellm_provider": "novita", @@ -47366,7 +47415,9 @@ "max_output_tokens": 20000, "max_tokens": 20000, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/google/gemma-3-27b-it": { "litellm_provider": "novita", @@ -47404,7 +47455,8 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_response_schema": true }, "novita/Sao10K/L3-8B-Stheno-v3.2": { "litellm_provider": "novita", @@ -47472,7 +47524,9 @@ "supports_parallel_function_calling": true, "supports_tool_choice": true, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "cache_read_input_token_cost": 2.5e-08, + "supports_prompt_caching": true }, "novita/qwen/qwen3-vl-30b-a3b-instruct": { "litellm_provider": "novita", @@ -47593,10 +47647,12 @@ "input_cost_per_token": 3e-08, "output_cost_per_token": 3e-08, "max_input_tokens": 128000, - "max_output_tokens": 20000, - "max_tokens": 20000, + "max_output_tokens": 8192, + "max_tokens": 8192, "supports_system_messages": true, - "supports_reasoning": true + "supports_reasoning": true, + "supports_function_calling": true, + "supports_tool_choice": true }, "novita/qwen/qwen2.5-7b-instruct": { "litellm_provider": "novita", @@ -47604,8 +47660,8 @@ "input_cost_per_token": 7e-08, "output_cost_per_token": 7e-08, "max_input_tokens": 32000, - "max_output_tokens": 32000, - "max_tokens": 32000, + "max_output_tokens": 8192, + "max_tokens": 8192, "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, @@ -50838,14 +50894,14 @@ "supports_embedding_image_input": true }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731": { - "cache_read_input_token_cost": 2.8e-08, - "input_cost_per_token": 1.4e-07, + "cache_read_input_token_cost": 7e-09, + "input_cost_per_token": 2.2e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", - "output_cost_per_token": 2.8e-07, + "output_cost_per_token": 6.6e-07, "source": "https://docs.fireworks.ai/serverless/pricing", "supports_function_calling": true, "supports_reasoning": true, @@ -51152,5 +51208,2015 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true + }, + "novita/zai-org/glm-5.3": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v4-pro-0813": { + "cache_read_input_token_cost": 1.3200000000000002e-07, + "input_cost_per_token": 1.32e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 3.96e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/moonshotai/kimi-k3": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 3e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 1048576, + "max_tokens": 1048576, + "mode": "chat", + "output_cost_per_token": 1.5e-05, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/tencent/hy3": { + "cache_read_input_token_cost": 3.5e-08, + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 5.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5.2": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/moonshotai/kimi-k2.7-code": { + "cache_read_input_token_cost": 1.9e-07, + "input_cost_per_token": 9.499999999999999e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash-vision-exp": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 4.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 1.32e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash-0731": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 4.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 1.32e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/mindai/macaron-v1-venti": { + "cache_read_input_token_cost": 3e-07, + "input_cost_per_token": 1.5e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.5e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m3": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek-v4-flash": { + "cache_read_input_token_cost": 2.8e-08, + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 2.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v4-pro": { + "cache_read_input_token_cost": 1.35e-07, + "input_cost_per_token": 1.6000000000000001e-06, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 393216, + "max_tokens": 393216, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/inclusionai/ling-3.0-flash-fast": { + "cache_read_input_token_cost": 1.2e-08, + "input_cost_per_token": 6e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.8-max": { + "cache_read_input_token_cost": 2.5e-07, + "input_cost_per_token": 2e-06, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 6e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/inclusionai/ling-3.0-flash": { + "cache_read_input_token_cost": 1.2e-08, + "input_cost_per_token": 6e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/mindai/macaron-v1-tall": { + "cache_read_input_token_cost": 8e-08, + "input_cost_per_token": 4.5000000000000003e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 2.6e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/stepfun/step-3.7-flash": { + "cache_read_input_token_cost": 4e-08, + "input_cost_per_token": 2.0000000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 256000, + "max_tokens": 256000, + "mode": "chat", + "output_cost_per_token": 1.15e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/nvidia/nemotron-3-nano-30b-a3b": { + "input_cost_per_token": 5.0000000000000004e-08, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 2.0000000000000002e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/baidu/cobuddy": { + "cache_read_input_token_cost": 7e-08, + "input_cost_per_token": 2.8e-07, + "litellm_provider": "novita", + "max_input_tokens": 131072, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.13e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/xiaomimimo/mimo-v2.5": { + "cache_read_input_token_cost": 3.4e-09, + "input_cost_per_token": 1.6800000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.3600000000000004e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.7-max": { + "cache_read_input_token_cost": 2.5e-07, + "input_cost_per_token": 1.25e-06, + "litellm_provider": "novita", + "max_input_tokens": 1000000, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.75e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/xiaomimimo/mimo-v2.5-pro": { + "cache_read_input_token_cost": 4.3e-09, + "input_cost_per_token": 5.22e-07, + "litellm_provider": "novita", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.044e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.6-27b": { + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.6000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/moonshotai/kimi-k2.6": { + "cache_read_input_token_cost": 1.6e-07, + "input_cost_per_token": 8.000000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 3.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-5.1": { + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.38e-06, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.7-highspeed": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5v-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/google/gemma-4-26b-a4b-it": { + "input_cost_per_token": 1.3e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/google/gemma-4-31b-it": { + "input_cost_per_token": 1.4e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-5-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "novita", + "max_input_tokens": 202800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.7": { + "cache_read_input_token_cost": 6e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/minimax/minimax-m2.5-highspeed": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131100, + "max_tokens": 131100, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.5-27b": { + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 2.4e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-122b-a10b": { + "input_cost_per_token": 4.0000000000000003e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-35b-a3b": { + "input_cost_per_token": 2.5e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/qwen/qwen3.5-397b-a17b": { + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 3.6000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/minimax/minimax-m2.5": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 3e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131100, + "max_tokens": 131100, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-5": { + "cache_read_input_token_cost": 2.0000000000000002e-07, + "input_cost_per_token": 1e-06, + "litellm_provider": "novita", + "max_input_tokens": 202800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.2000000000000003e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3-coder-next": { + "input_cost_per_token": 2.0000000000000002e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.5e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-ocr-2": { + "input_cost_per_token": 3e-08, + "litellm_provider": "novita", + "max_input_tokens": 8192, + "max_output_tokens": 8192, + "max_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 3e-08, + "source": "https://novita.ai/pricing", + "supports_vision": true + }, + "novita/moonshotai/kimi-k2.5": { + "cache_read_input_token_cost": 1.0000000000000001e-07, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 3e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/zai-org/glm-4.7-h": { + "cache_read_input_token_cost": 1.1e-07, + "input_cost_per_token": 6e-07, + "litellm_provider": "novita", + "max_input_tokens": 204800, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 2.2e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/zai-org/glm-4.7-flash": { + "cache_read_input_token_cost": 1e-08, + "input_cost_per_token": 7e-08, + "litellm_provider": "novita", + "max_input_tokens": 200000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.0000000000000003e-07, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/qwen/qwen3.6-35b-a3b": { + "input_cost_per_token": 2.48e-07, + "litellm_provider": "novita", + "max_input_tokens": 262144, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.4850000000000002e-06, + "source": "https://novita.ai/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "novita/deepseek/deepseek_v3": { + "input_cost_per_token": 8.900000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 16000, + "max_tokens": 16000, + "mode": "chat", + "output_cost_per_token": 8.900000000000001e-07, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-r1": { + "input_cost_per_token": 4e-06, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 16000, + "max_tokens": 16000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-v3/community": { + "input_cost_per_token": 8.900000000000001e-07, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 8000, + "max_tokens": 8000, + "mode": "chat", + "output_cost_per_token": 8.900000000000001e-07, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/deepseek/deepseek-r1/community": { + "input_cost_per_token": 4e-06, + "litellm_provider": "novita", + "max_input_tokens": 64000, + "max_output_tokens": 8000, + "max_tokens": 8000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/thudm/glm-4-32b-0414": { + "input_cost_per_token": 5.5e-07, + "litellm_provider": "novita", + "max_input_tokens": 32000, + "max_output_tokens": 32000, + "max_tokens": 32000, + "mode": "chat", + "output_cost_per_token": 1.66e-06, + "source": "https://api.novita.ai/v3/openai/models", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, + "novita/meta-llama/llama-3.2-1b-instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "novita", + "max_input_tokens": 131000, + "max_output_tokens": 32000, + "max_tokens": 32000, + "mode": "chat", + "output_cost_per_token": 2e-08, + "source": "https://api.novita.ai/v3/openai/models", + "supports_response_schema": true, + "supports_vision": false + }, + "wandb/deepseek-ai/DeepSeek-V4-Flash": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 2.8e-07, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/deepseek-ai/DeepSeek-V4-Flash-0731": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.3e-07, + "output_cost_per_token": 2.8e-07, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/deepseek-ai/DeepSeek-V4-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.15e-06, + "output_cost_per_token": 2.55e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/google/gemma-4-31B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3.4e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/ibm-granite/granite-4.1-8b": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/JetBrains/Mellum2-12B-A2.5B-Instruct": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/meta-llama/Llama-3.1-70B-Instruct": { + "max_tokens": 128000, + "max_input_tokens": 128000, + "input_cost_per_token": 8e-07, + "output_cost_per_token": 8e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/MiniMaxAI/MiniMax-M3": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.3e-07, + "output_cost_per_token": 9.6e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/moonshotai/Kimi-K2.7-Code": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.1e-07, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/moonshotai/Kimi-K2.6": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6.5e-07, + "output_cost_per_token": 3.41e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 2.5e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 2.75e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/OpenPipe/Qwen3-14B-Instruct": { + "max_tokens": 32768, + "max_input_tokens": 32768, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 2.2e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.8-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.6-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.25e-06, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.6-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6e-07, + "output_cost_per_token": 3.6e-06, + "cache_read_input_token_cost": 1.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3.5-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.25e-06, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": true, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/Qwen/Qwen3-30B-A3B-Instruct-2507": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 3e-07, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "wandb/zai-org/GLM-5.2": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.6e-07, + "output_cost_per_token": 2.42e-06, + "cache_read_input_token_cost": 1.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "wandb", + "mode": "chat", + "supports_vision": false, + "source": "https://wandb.ai/site/pricing/tokens/" + }, + "deepinfra/openai/gpt-oss-120b-Turbo": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 1.5e-07, + "output_cost_per_token": 6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M2.7": { + "max_tokens": 196608, + "max_input_tokens": 196608, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it-Ultra": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2.7e-07, + "output_cost_per_token": 7.6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.5": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 2.25e-06, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.7-Flash": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 6e-08, + "output_cost_per_token": 4e-07, + "cache_read_input_token_cost": 1e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.6": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-4-8": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-sonnet-4-6": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 3e-06, + "output_cost_per_token": 1.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.5-flash": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 1.5e-06, + "output_cost_per_token": 9e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/XiaomiMiMo/MiMo-V2.5": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it-turbo": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 3.4e-07, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/thinkingmachines/Inkling-Small": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 1.2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/meta-models/Muse-Glimmer-30B": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 3e-07, + "output_cost_per_token": 1.2e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-Max-Thinking": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-VL-235B-A22B-Instruct": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 8.8e-07, + "cache_read_input_token_cost": 1.1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3-VL-30B-A3B-Instruct": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.5e-07, + "output_cost_per_token": 6e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.6e-07, + "output_cost_per_token": 2.6e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.6-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 9.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/Nemotron-Content-Safety-3.5": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 2e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/thinkingmachines/Inkling": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 9.5e-07, + "output_cost_per_token": 4.05e-06, + "cache_read_input_token_cost": 1.6e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.6": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 1.5e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Pro-0813": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.3e-06, + "output_cost_per_token": 2.6e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.7-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 2.5e-06, + "output_cost_per_token": 7.5e-06, + "cache_read_input_token_cost": 5e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-mini": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 4e-07, + "cache_read_input_token_cost": 2e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-2.4T-A95B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 6e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M3": { + "max_tokens": 524288, + "max_input_tokens": 524288, + "input_cost_per_token": 2.8e-07, + "output_cost_per_token": 1.1e-06, + "cache_read_input_token_cost": 5.6e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.1-flash-lite": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 1.5e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.7-flash": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 3.75e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/inclusionAI/Ling-3.0-flash": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 6e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.2e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/stepfun-ai/Step-3.7-Flash": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 1.15e-06, + "cache_read_input_token_cost": 4e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-35B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 1e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-1.8": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 2.5e-07, + "output_cost_per_token": 2e-06, + "cache_read_input_token_cost": 5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/tencent/Hy3": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.4e-07, + "output_cost_per_token": 5.8e-07, + "cache_read_input_token_cost": 3.5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-code": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/ByteDance/Seed-2.0-pro": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 6e-07, + "output_cost_per_token": 2.08e-06, + "cache_read_input_token_cost": 1.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/Nemotron-3-Nano-30B-A3B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 5e-08, + "output_cost_per_token": 2e-07, + "cache_read_input_token_cost": 2.5e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K2.7-Code": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 6.8e-07, + "output_cost_per_token": 3.4e-06, + "cache_read_input_token_cost": 1.36e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-sonnet-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 1e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-397B-A17B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 4.5e-07, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 2.2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Flash-0731": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 8e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.6e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-E4B-it": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-08, + "output_cost_per_token": 1e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V3.2": { + "max_tokens": 163840, + "max_input_tokens": 163840, + "input_cost_per_token": 2.6e-07, + "output_cost_per_token": 3.8e-07, + "cache_read_input_token_cost": 1.3e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.8-Max": { + "max_tokens": 256000, + "max_input_tokens": 256000, + "input_cost_per_token": 1.65e-06, + "output_cost_per_token": 4.951e-06, + "cache_read_input_token_cost": 2.06e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-fable-5": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 1e-05, + "output_cost_per_token": 5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 5e-07, + "output_cost_per_token": 2.2e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-122B-A10B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 2.9e-07, + "output_cost_per_token": 2.4e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5.1": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 1.05e-06, + "output_cost_per_token": 3.5e-06, + "cache_read_input_token_cost": 2.05e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1.3e-06, + "output_cost_per_token": 2.6e-06, + "cache_read_input_token_cost": 1e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 8.5e-08, + "output_cost_per_token": 4e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-5.2": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 7.5e-07, + "output_cost_per_token": 2.4e-06, + "cache_read_input_token_cost": 1.4e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/moonshotai/Kimi-K3": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 2.85e-06, + "output_cost_per_token": 1.425e-05, + "cache_read_input_token_cost": 2.85e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-opus-4-7": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 5e-06, + "output_cost_per_token": 2.5e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.6-27B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 3.2e-07, + "output_cost_per_token": 3.2e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-26B-A4B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 7e-08, + "output_cost_per_token": 3.4e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemini-3.1-pro": { + "max_tokens": 1000000, + "max_input_tokens": 1000000, + "input_cost_per_token": 2e-06, + "output_cost_per_token": 1.2e-05, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/XiaomiMiMo/MiMo-V2.5-Pro": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 3e-06, + "cache_read_input_token_cost": 2e-07, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/anthropic/claude-haiku-4-5": { + "max_tokens": 200000, + "max_input_tokens": 200000, + "input_cost_per_token": 1e-06, + "output_cost_per_token": 5e-06, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/deepseek-ai/DeepSeek-V4-Flash": { + "max_tokens": 1048576, + "max_input_tokens": 1048576, + "input_cost_per_token": 9e-08, + "output_cost_per_token": 1.8e-07, + "cache_read_input_token_cost": 1.8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/openai/gpt-oss-120b-Ultra": { + "max_tokens": 131072, + "max_input_tokens": 131072, + "input_cost_per_token": 2e-07, + "output_cost_per_token": 9.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/Qwen/Qwen3.5-9B": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1e-07, + "output_cost_per_token": 1.5e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/MiniMaxAI/MiniMax-M2.7-Turbo": { + "max_tokens": 196608, + "max_input_tokens": 196608, + "input_cost_per_token": 3.8e-07, + "output_cost_per_token": 1.7e-06, + "cache_read_input_token_cost": 7e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/zai-org/GLM-4.7": { + "max_tokens": 202752, + "max_input_tokens": 202752, + "input_cost_per_token": 4e-07, + "output_cost_per_token": 1.75e-06, + "cache_read_input_token_cost": 8e-08, + "supports_prompt_caching": true, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": false, + "source": "https://deepinfra.com/pricing" + }, + "deepinfra/google/gemma-4-31B-it": { + "max_tokens": 262144, + "max_input_tokens": 262144, + "input_cost_per_token": 1.3e-07, + "output_cost_per_token": 3.8e-07, + "litellm_provider": "deepinfra", + "mode": "chat", + "supports_function_calling": true, + "supports_tool_choice": true, + "supports_response_schema": true, + "supports_reasoning": true, + "supports_vision": true, + "source": "https://deepinfra.com/pricing" } } From 2a5e071cab9a4426fdceced11c1f9702f46d4345 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 25 Aug 2026 13:46:36 -0700 Subject: [PATCH 10/44] feat(together_ai): map reasoning_effort per model class --- .../llms/together_ai/chat/transformation.py | 70 ++++++++- .../test_together_ai_chat_transformation.py | 140 ++++++++++++++++-- 2 files changed, 197 insertions(+), 13 deletions(-) diff --git a/litellm/llms/together_ai/chat/transformation.py b/litellm/llms/together_ai/chat/transformation.py index 88fd79f2366..ee6a813490e 100644 --- a/litellm/llms/together_ai/chat/transformation.py +++ b/litellm/llms/together_ai/chat/transformation.py @@ -4,17 +4,77 @@ Translates from OpenAI's `/v1/chat/completions` to Together AI's `/v1/chat/compl Docs: https://docs.together.ai/docs/chat-overview """ +from collections.abc import Mapping from types import MappingProxyType from typing import Final +from typing_extensions import ReadOnly, TypedDict + from litellm._logging import verbose_logger -from litellm.utils import supports_function_calling +from litellm.utils import supports_function_calling, supports_reasoning from ...openai.chat.gpt_transformation import OpenAIGPTConfig FUNCTION_CALLING_ONLY_PARAMS: Final = ("tools", "tool_choice", "function_call", "response_format") PLAIN_TEXT_RESPONSE_FORMAT: Final = MappingProxyType({"type": "text"}) +ADJUSTABLE_EFFORT_REASONING_MODELS: Final = frozenset( + { + "openai/gpt-oss-120b", + "openai/gpt-oss-20b", + } +) +HYBRID_REASONING_MODELS: Final = frozenset( + { + "MiniMaxAI/MiniMax-M3", + "Qwen/Qwen3.5-9B", + "Qwen/Qwen3.6-Plus", + "deepseek-ai/DeepSeek-V4-Pro", + "moonshotai/Kimi-K3", + "nvidia/nemotron-3-ultra-550b-a55b", + "zai-org/GLM-5.2", + } +) +HIGH_MAX_EFFORT_MODEL_PREFIX: Final = "deepseek-ai/DeepSeek-V4-Pro" +EFFORT_TRANSLATION: Final = MappingProxyType({"minimal": "low", "xhigh": "high", "max": "high"}) +HIGH_MAX_EFFORT_TRANSLATION: Final = MappingProxyType( + {"minimal": "high", "low": "high", "medium": "high", "high": "max", "xhigh": "max"} +) + + +class TogetherReasoningToggle(TypedDict): + enabled: ReadOnly[bool] + + +def _supports_together_reasoning(model: str) -> bool: + if model in ADJUSTABLE_EFFORT_REASONING_MODELS or model in HYBRID_REASONING_MODELS: + return True + if model.startswith(HIGH_MAX_EFFORT_MODEL_PREFIX): + return True + return supports_reasoning(model, custom_llm_provider="together_ai") + + +def _adjustable_effort(effort: str, model: str) -> str: + if effort == "none": + verbose_logger.debug( + "together_ai model %s cannot disable reasoning; mapping reasoning_effort=none to low", model + ) + return "low" + return EFFORT_TRANSLATION.get(effort, effort) + + +def _reasoning_effort_payload(effort: str, model: str) -> Mapping[str, object]: + if effort == "default": + return MappingProxyType({}) + if model in ADJUSTABLE_EFFORT_REASONING_MODELS: + return MappingProxyType({"reasoning_effort": _adjustable_effort(effort, model)}) + if effort == "none": + disable_reasoning: Final[TogetherReasoningToggle] = {"enabled": False} + return MappingProxyType({"reasoning": disable_reasoning}) + if model.startswith(HIGH_MAX_EFFORT_MODEL_PREFIX): + return MappingProxyType({"reasoning_effort": HIGH_MAX_EFFORT_TRANSLATION.get(effort, effort)}) + return MappingProxyType({"reasoning_effort": EFFORT_TRANSLATION.get(effort, effort)}) + class TogetherAIChatConfig(OpenAIGPTConfig): def get_supported_openai_params(self, model: str) -> list: @@ -25,6 +85,8 @@ class TogetherAIChatConfig(OpenAIGPTConfig): verbose_logger.debug("Error getting supported openai params: %s", e) supported_params: Final = super().get_supported_openai_params(model) + if _supports_together_reasoning(model): + supported_params.append("reasoning_effort") if supports_fc is True: return supported_params verbose_logger.debug( @@ -45,4 +107,10 @@ class TogetherAIChatConfig(OpenAIGPTConfig): if mapped_openai_params.get("response_format") == PLAIN_TEXT_RESPONSE_FORMAT: mapped_openai_params.pop("response_format") + effort: Final = mapped_openai_params.get("reasoning_effort") + if not isinstance(effort, str): + return mapped_openai_params + mapped_openai_params.pop("reasoning_effort") + for key, value in _reasoning_effort_payload(effort, model).items(): + mapped_openai_params.setdefault(key, value) return mapped_openai_params diff --git a/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py b/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py index 6216d3bf225..e99bbc46523 100644 --- a/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py +++ b/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py @@ -16,10 +16,24 @@ TOOL_CALLING_MODEL = "openai/gpt-oss-20b" REASONING_MODEL = "deepseek-ai/DeepSeek-V3.1" PLAIN_MODEL = "Qwen/Qwen3-235B-A22B-fp8-tput" UNMAPPED_MODEL = "MiniMaxAI/MiniMax-M3" +ADJUSTABLE_REASONING_MODEL = "openai/gpt-oss-120b" +HYBRID_REASONING_MODEL = "Qwen/Qwen3.5-9B" +HIGH_MAX_REASONING_MODEL = "deepseek-ai/DeepSeek-V4-Pro" +REGISTRY_FLAGGED_REASONING_MODEL = "zai-org/GLM-4.6" +NON_REASONING_MODEL = "meta-llama/Llama-3.3-70B-Instruct-Turbo" FUNCTION_CALLING_PARAMS = ("tools", "tool_choice", "function_call", "response_format") +def _map_reasoning_effort(model: str, effort: str) -> dict: + return TogetherAIChatConfig().map_openai_params( + non_default_params={"reasoning_effort": effort}, + optional_params={}, + model=model, + drop_params=False, + ) + + @pytest.fixture(autouse=True) def force_local_model_cost(monkeypatch): monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") @@ -103,6 +117,116 @@ def test_map_openai_params_keeps_json_response_format(): assert mapped["response_format"] == response_format +@pytest.mark.parametrize( + "model", + [ADJUSTABLE_REASONING_MODEL, HYBRID_REASONING_MODEL, HIGH_MAX_REASONING_MODEL, REGISTRY_FLAGGED_REASONING_MODEL], +) +def test_supported_params_includes_reasoning_effort_for_reasoning_models(model): + supported = TogetherAIChatConfig().get_supported_openai_params(model=model) + + assert "reasoning_effort" in supported + + +@pytest.mark.parametrize("model", [NON_REASONING_MODEL, PLAIN_MODEL]) +def test_supported_params_excludes_reasoning_effort_for_non_reasoning_models(model): + supported = TogetherAIChatConfig().get_supported_openai_params(model=model) + + assert "reasoning_effort" not in supported + + +@pytest.mark.parametrize( + "effort, expected", + [("low", "low"), ("medium", "medium"), ("high", "high"), ("minimal", "low"), ("xhigh", "high"), ("max", "high")], +) +def test_adjustable_model_translates_reasoning_effort(effort, expected): + mapped = _map_reasoning_effort(ADJUSTABLE_REASONING_MODEL, effort) + + assert mapped["reasoning_effort"] == expected + assert "reasoning" not in mapped + + +def test_adjustable_model_cannot_disable_reasoning_so_none_becomes_low(): + mapped = _map_reasoning_effort(ADJUSTABLE_REASONING_MODEL, "none") + + assert mapped["reasoning_effort"] == "low" + assert "reasoning" not in mapped + + +@pytest.mark.parametrize( + "effort, expected", + [("low", "low"), ("medium", "medium"), ("high", "high"), ("minimal", "low"), ("xhigh", "high"), ("max", "high")], +) +def test_hybrid_model_translates_reasoning_effort(effort, expected): + mapped = _map_reasoning_effort(HYBRID_REASONING_MODEL, effort) + + assert mapped["reasoning_effort"] == expected + assert "reasoning" not in mapped + + +@pytest.mark.parametrize("model", [HYBRID_REASONING_MODEL, HIGH_MAX_REASONING_MODEL, REGISTRY_FLAGGED_REASONING_MODEL]) +def test_reasoning_effort_none_becomes_reasoning_toggle(model): + mapped = _map_reasoning_effort(model, "none") + + assert mapped["reasoning"] == {"enabled": False} + assert "reasoning_effort" not in mapped + + +def test_reasoning_effort_none_does_not_clobber_user_reasoning(): + mapped = TogetherAIChatConfig().map_openai_params( + non_default_params={"reasoning_effort": "none"}, + optional_params={"reasoning": {"enabled": True}}, + model=HYBRID_REASONING_MODEL, + drop_params=False, + ) + + assert mapped["reasoning"] == {"enabled": True} + assert "reasoning_effort" not in mapped + + +@pytest.mark.parametrize( + "effort, expected", + [("minimal", "high"), ("low", "high"), ("medium", "high"), ("high", "max"), ("xhigh", "max"), ("max", "max")], +) +def test_deepseek_v4_pro_remaps_to_high_max(effort, expected): + mapped = _map_reasoning_effort(HIGH_MAX_REASONING_MODEL, effort) + + assert mapped["reasoning_effort"] == expected + + +def test_deepseek_v4_pro_dated_variant_remaps_via_prefix(): + mapped = _map_reasoning_effort(f"{HIGH_MAX_REASONING_MODEL}-0813", "low") + + assert mapped["reasoning_effort"] == "high" + + +@pytest.mark.parametrize("model", [ADJUSTABLE_REASONING_MODEL, HYBRID_REASONING_MODEL, HIGH_MAX_REASONING_MODEL]) +def test_reasoning_effort_default_is_dropped(model): + mapped = _map_reasoning_effort(model, "default") + + assert "reasoning_effort" not in mapped + assert "reasoning" not in mapped + + +def test_get_optional_params_translates_reasoning_effort_for_together(): + optional_params = litellm.get_optional_params( + model=ADJUSTABLE_REASONING_MODEL, + custom_llm_provider="together_ai", + reasoning_effort="max", + ) + + assert optional_params["reasoning_effort"] == "high" + + +def test_get_optional_params_rejects_reasoning_effort_for_non_reasoning_together_model(): + with pytest.raises(litellm.UnsupportedParamsError): + litellm.get_optional_params( + model=NON_REASONING_MODEL, + custom_llm_provider="together_ai", + reasoning_effort="low", + drop_params=False, + ) + + def _transform_response(message: dict) -> ModelResponse: raw_response_json = { "id": "chatcmpl-test", @@ -136,26 +260,20 @@ def _transform_response(message: dict) -> ModelResponse: def test_transform_response_maps_reasoning_to_reasoning_content(): - result = _transform_response( - {"role": "assistant", "content": "4", "reasoning": "2+2 equals 4"} - ) + result = _transform_response({"role": "assistant", "content": "4", "reasoning": "2+2 equals 4"}) assert result.choices[0].message.content == "4" assert result.choices[0].message.reasoning_content == "2+2 equals 4" def test_transform_response_preserves_reasoning_content_field(): - result = _transform_response( - {"role": "assistant", "content": "4", "reasoning_content": "adding 2 and 2"} - ) + result = _transform_response({"role": "assistant", "content": "4", "reasoning_content": "adding 2 and 2"}) assert result.choices[0].message.reasoning_content == "adding 2 and 2" def test_streaming_chunk_maps_delta_reasoning_to_reasoning_content(): - iterator = TogetherAIChatConfig().get_model_response_iterator( - streaming_response=iter(()), sync_stream=True - ) + iterator = TogetherAIChatConfig().get_model_response_iterator(streaming_response=iter(()), sync_stream=True) assert isinstance(iterator, OpenAIChatCompletionStreamingHandler) parsed = iterator.chunk_parser( @@ -179,9 +297,7 @@ def test_together_ai_config_alias_points_at_chat_config(): def test_provider_config_manager_returns_together_chat_config(): from litellm.utils import ProviderConfigManager - config = ProviderConfigManager.get_provider_chat_config( - model=REASONING_MODEL, provider=LlmProviders.TOGETHER_AI - ) + config = ProviderConfigManager.get_provider_chat_config(model=REASONING_MODEL, provider=LlmProviders.TOGETHER_AI) assert isinstance(config, TogetherAIChatConfig) From 6fafb46731cd94881e715c685d983c7b817fdf61 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 25 Aug 2026 16:45:50 -0700 Subject: [PATCH 11/44] fix(cost): apply Together AI cache read pricing and per-model registry rates --- litellm/cost_calculator.py | 11 +++- litellm/llms/together_ai/cost_calculator.py | 6 ++ ...odel_prices_and_context_window_backup.json | 38 +++++++++++-- model_prices_and_context_window.json | 38 +++++++++++-- tests/test_litellm/test_cost_calculator.py | 57 +++++++++++++++++++ .../test_together_ai_model_metadata.py | 48 ++++++++++++++++ 6 files changed, 183 insertions(+), 15 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 6536941a094..6b1218415af 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -75,7 +75,10 @@ from litellm.llms.perplexity.cost_calculator import ( from litellm.llms.tencent.cost_calculator import ( cost_per_token as tencent_cost_per_token, ) -from litellm.llms.together_ai.cost_calculator import get_model_params_and_category +from litellm.llms.together_ai.cost_calculator import ( + get_model_params_and_category, + has_together_registry_entry, +) from litellm.llms.vertex_ai.cost_calculator import ( cost_per_character as google_cost_per_character, ) @@ -1551,8 +1554,10 @@ def completion_cost( return MCPCostCalculator.calculate_mcp_tool_call_cost(litellm_logging_obj=litellm_logging_obj) # Calculate cost based on prompt_tokens, completion_tokens - if "togethercomputer" in model or "together_ai" in model or custom_llm_provider == "together_ai": - # together ai prices based on size of llm + if ( + "togethercomputer" in model or "together_ai" in model or custom_llm_provider == "together_ai" + ) and not has_together_registry_entry(model, litellm.model_cost): + # together ai prices unmapped models based on size of llm # get_model_params_and_category takes a model name and returns the category of LLM size it is in model_prices_and_context_window.json model = get_model_params_and_category(model, call_type=CallTypes(call_type)) diff --git a/litellm/llms/together_ai/cost_calculator.py b/litellm/llms/together_ai/cost_calculator.py index 431e94f1442..2f492496e7b 100644 --- a/litellm/llms/together_ai/cost_calculator.py +++ b/litellm/llms/together_ai/cost_calculator.py @@ -3,6 +3,7 @@ Handles calculating cost for together ai models """ import re +from collections.abc import Mapping from typing import Final from litellm.constants import ( @@ -18,6 +19,11 @@ from litellm.constants import ( from litellm.types.utils import CallTypes +def has_together_registry_entry(model: str, cost_map: Mapping[str, object]) -> bool: + stripped: Final = model.removeprefix("together_ai/") + return f"together_ai/{stripped}" in cost_map + + # Extract the number of billion parameters from the model name # only used for together_computer LLMs def get_model_params_and_category(model_name, call_type: CallTypes) -> str: diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 9ca8d9e1bac..528f428c802 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -38309,6 +38309,7 @@ "supports_tool_choice": true }, "together_ai/MiniMaxAI/MiniMax-M3": { + "cache_read_input_token_cost": 6e-08, "input_cost_per_token": 3e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38319,6 +38320,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -38362,14 +38364,16 @@ "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { - "input_cost_per_token": 1.25e-06, + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, "mode": "chat", - "output_cost_per_token": 3.75e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "output_cost_per_token": 7.5e-06, + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { "input_cost_per_token": 3.2e-07, @@ -38382,6 +38386,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { + "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1010000, @@ -38389,7 +38394,8 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6.25e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { "input_cost_per_token": 1e-07, @@ -38402,6 +38408,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { + "cache_read_input_token_cost": 3e-08, "input_cost_per_token": 1.4e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38412,10 +38419,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", "max_input_tokens": 512000, @@ -38426,11 +38435,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.32e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38441,6 +38452,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, @@ -38490,6 +38502,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-models/Muse-Glimmer-30B": { + "cache_read_input_token_cost": 4e-08, "input_cost_per_token": 3.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 131072, @@ -38497,9 +38510,11 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -38510,11 +38525,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/moonshotai/Kimi-K3": { + "cache_read_input_token_cost": 3e-07, "input_cost_per_token": 3e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38525,12 +38542,14 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", "max_input_tokens": 512288, @@ -38541,6 +38560,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true @@ -38556,6 +38576,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/thinkingmachines/Inkling": { + "cache_read_input_token_cost": 1.7e-07, "input_cost_per_token": 1e-06, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38566,10 +38587,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/thinkingmachines/Inkling-Small": { + "cache_read_input_token_cost": 1e-07, "input_cost_per_token": 5e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38577,9 +38600,11 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { + "cache_read_input_token_cost": 2.6e-07, "input_cost_per_token": 1.4e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048575, @@ -38590,6 +38615,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 9ca8d9e1bac..528f428c802 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -38309,6 +38309,7 @@ "supports_tool_choice": true }, "together_ai/MiniMaxAI/MiniMax-M3": { + "cache_read_input_token_cost": 6e-08, "input_cost_per_token": 3e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38319,6 +38320,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -38362,14 +38364,16 @@ "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { - "input_cost_per_token": 1.25e-06, + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, "mode": "chat", - "output_cost_per_token": 3.75e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "output_cost_per_token": 7.5e-06, + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { "input_cost_per_token": 3.2e-07, @@ -38382,6 +38386,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { + "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1010000, @@ -38389,7 +38394,8 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6.25e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { "input_cost_per_token": 1e-07, @@ -38402,6 +38408,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { + "cache_read_input_token_cost": 3e-08, "input_cost_per_token": 1.4e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38412,10 +38419,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", "max_input_tokens": 512000, @@ -38426,11 +38435,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.32e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38441,6 +38452,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, @@ -38490,6 +38502,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-models/Muse-Glimmer-30B": { + "cache_read_input_token_cost": 4e-08, "input_cost_per_token": 3.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 131072, @@ -38497,9 +38510,11 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -38510,11 +38525,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/moonshotai/Kimi-K3": { + "cache_read_input_token_cost": 3e-07, "input_cost_per_token": 3e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38525,12 +38542,14 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", "max_input_tokens": 512288, @@ -38541,6 +38560,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true @@ -38556,6 +38576,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/thinkingmachines/Inkling": { + "cache_read_input_token_cost": 1.7e-07, "input_cost_per_token": 1e-06, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38566,10 +38587,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/thinkingmachines/Inkling-Small": { + "cache_read_input_token_cost": 1e-07, "input_cost_per_token": 5e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38577,9 +38600,11 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { + "cache_read_input_token_cost": 2.6e-07, "input_cost_per_token": 1.4e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048575, @@ -38590,6 +38615,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py index 1f1c9be973f..bf230bf794c 100644 --- a/tests/test_litellm/test_cost_calculator.py +++ b/tests/test_litellm/test_cost_calculator.py @@ -3781,3 +3781,60 @@ def test_completion_cost_prices_anthropic_shaped_cache_read_tokens(_local_model_ ) assert cost == pytest.approx(3 * 4e-6 + 4014 * 4e-7 + 5 * 2e-5, rel=1e-9) + + +def _together_chat_response(model: str, prompt_tokens: int, completion_tokens: int, cached_tokens: int) -> ModelResponse: + return ModelResponse( + id="chatcmpl-together-cache", + choices=[{"finish_reason": "stop", "index": 0, "message": {"content": "acknowledged", "role": "assistant"}}], + created=1756164000, + model=model, + object="chat.completion", + usage=Usage( + prompt_tokens=prompt_tokens, + completion_tokens=completion_tokens, + total_tokens=prompt_tokens + completion_tokens, + prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=cached_tokens), + ), + ) + + +def test_completion_cost_prices_together_cached_tokens_at_cache_read_rate(_local_model_cost_map): + """Regression: Together reports prompt_tokens_details.cached_tokens but no together_ai + registry entry carried cache_read_input_token_cost, so cache-hit tokens were priced at + 0.0 and spend on cache-heavy workloads was understated.""" + + cost = completion_cost( + completion_response=_together_chat_response( + model="deepseek-ai/DeepSeek-V4-Flash-0731", prompt_tokens=7864, completion_tokens=16, cached_tokens=7863 + ), + custom_llm_provider="together_ai", + ) + + assert cost == pytest.approx(1 * 1.4e-07 + 7863 * 3e-08 + 16 * 2.8e-07, rel=1e-9) + + +def test_completion_cost_together_mapped_model_skips_size_bucket(_local_model_cost_map): + """Regression: any together model whose name matches (\\d+b) was rewritten to a + together-ai-* size bucket before the registry lookup, so mapped models like + Muse-Glimmer-30B never used their per-model rates, cache fields included.""" + + cost = completion_cost( + completion_response=_together_chat_response( + model="meta-models/Muse-Glimmer-30B", prompt_tokens=63, completion_tokens=16, cached_tokens=0 + ), + custom_llm_provider="together_ai", + ) + + assert cost == pytest.approx(63 * 3.5e-07 + 16 * 1.5e-06, rel=1e-9) + + +def test_completion_cost_together_unmapped_model_still_uses_size_bucket(_local_model_cost_map): + cost = completion_cost( + completion_response=_together_chat_response( + model="qwen/Qwen2-72B-Instruct", prompt_tokens=23, completion_tokens=15, cached_tokens=0 + ), + custom_llm_provider="together_ai", + ) + + assert cost == pytest.approx((23 + 15) * 9e-07, rel=1e-9) diff --git a/tests/test_litellm/test_together_ai_model_metadata.py b/tests/test_litellm/test_together_ai_model_metadata.py index 5a0aadf4737..e4531a1c283 100644 --- a/tests/test_litellm/test_together_ai_model_metadata.py +++ b/tests/test_litellm/test_together_ai_model_metadata.py @@ -159,3 +159,51 @@ def test_together_backup_cost_map_in_sync(cost_map: CostMap): together_main = {k: v for k, v in cost_map.items() if k.startswith("together_ai/")} together_backup = {k: v for k, v in backup.items() if k.startswith("together_ai/")} assert together_backup == together_main + + +CACHED_INPUT_MODELS: Final = ( + "together_ai/moonshotai/Kimi-K3", + "together_ai/zai-org/GLM-5.2", + "together_ai/meta-models/Muse-Glimmer-30B", + "together_ai/Qwen/Qwen3.8-2.4T-A95B", + "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813", + "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731", + "together_ai/thinkingmachines/Inkling", + "together_ai/MiniMaxAI/MiniMax-M3", + "together_ai/thinkingmachines/Inkling-Small", + "together_ai/moonshotai/Kimi-K2.7-Code", + "together_ai/deepseek-ai/DeepSeek-V4-Pro", + "together_ai/nvidia/nemotron-3-ultra-550b-a55b", + "together_ai/Qwen/Qwen3.7-Max", +) + + +@pytest.mark.parametrize("model", CACHED_INPUT_MODELS) +def test_together_cached_input_model_carries_cache_read_pricing(cost_map: CostMap, model: str): + info = cost_map.get(model) + assert info is not None, f"{model} missing from model_prices_and_context_window.json" + assert info.get("supports_prompt_caching") is True + cache_read = info.get("cache_read_input_token_cost") + assert isinstance(cache_read, float) + assert 0 < cache_read < info["input_cost_per_token"] + assert "cache_creation_input_token_cost" not in info + + +def test_together_prompt_caching_flag_implies_cache_read_rate(cost_map: CostMap): + for model, info in cost_map.items(): + if model.startswith("together_ai/") and info.get("supports_prompt_caching"): + assert "cache_read_input_token_cost" in info, f"{model} flags caching without a cache read rate" + + +def test_together_deepseek_v4_flash_cache_read_rate(cost_map: CostMap): + info = cost_map["together_ai/deepseek-ai/DeepSeek-V4-Flash-0731"] + assert info["input_cost_per_token"] == 1.4e-07 + assert info["cache_read_input_token_cost"] == 3e-08 + assert info["output_cost_per_token"] == 2.8e-07 + + +def test_together_qwen_37_max_repriced_to_current_together_rate(cost_map: CostMap): + info = cost_map["together_ai/Qwen/Qwen3.7-Max"] + assert info["input_cost_per_token"] == 2.5e-06 + assert info["output_cost_per_token"] == 7.5e-06 + assert info["cache_read_input_token_cost"] == 5e-07 From ac866e98c8d1d3a6622663f8625dd1cd6e05ea5a Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 25 Aug 2026 16:56:17 -0700 Subject: [PATCH 12/44] chore(cost): drop redundant together fallback comment --- litellm/cost_calculator.py | 3 --- 1 file changed, 3 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 6b1218415af..78a555940c8 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -1557,9 +1557,6 @@ def completion_cost( if ( "togethercomputer" in model or "together_ai" in model or custom_llm_provider == "together_ai" ) and not has_together_registry_entry(model, litellm.model_cost): - # together ai prices unmapped models based on size of llm - # get_model_params_and_category takes a model name and returns the category of LLM size it is in model_prices_and_context_window.json - model = get_model_params_and_category(model, call_type=CallTypes(call_type)) # replicate llms are calculate based on time for request running From abe9af622bbb6c6e497022876b3d0bbb26de9fcb Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:09:12 -0700 Subject: [PATCH 13/44] fix(cost): keep size buckets for Together registry rows without pricing --- litellm/cost_calculator.py | 4 ++-- litellm/llms/together_ai/cost_calculator.py | 5 +++-- tests/test_litellm/test_cost_calculator.py | 13 +++++++++++++ 3 files changed, 18 insertions(+), 4 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 78a555940c8..153b8457e38 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -77,7 +77,7 @@ from litellm.llms.tencent.cost_calculator import ( ) from litellm.llms.together_ai.cost_calculator import ( get_model_params_and_category, - has_together_registry_entry, + has_together_registry_pricing, ) from litellm.llms.vertex_ai.cost_calculator import ( cost_per_character as google_cost_per_character, @@ -1556,7 +1556,7 @@ def completion_cost( # Calculate cost based on prompt_tokens, completion_tokens if ( "togethercomputer" in model or "together_ai" in model or custom_llm_provider == "together_ai" - ) and not has_together_registry_entry(model, litellm.model_cost): + ) and not has_together_registry_pricing(model, litellm.model_cost): model = get_model_params_and_category(model, call_type=CallTypes(call_type)) # replicate llms are calculate based on time for request running diff --git a/litellm/llms/together_ai/cost_calculator.py b/litellm/llms/together_ai/cost_calculator.py index 2f492496e7b..6fc2c949fa6 100644 --- a/litellm/llms/together_ai/cost_calculator.py +++ b/litellm/llms/together_ai/cost_calculator.py @@ -19,9 +19,10 @@ from litellm.constants import ( from litellm.types.utils import CallTypes -def has_together_registry_entry(model: str, cost_map: Mapping[str, object]) -> bool: +def has_together_registry_pricing(model: str, cost_map: Mapping[str, object]) -> bool: stripped: Final = model.removeprefix("together_ai/") - return f"together_ai/{stripped}" in cost_map + entry: Final = cost_map.get(f"together_ai/{stripped}") + return isinstance(entry, Mapping) and "input_cost_per_token" in entry # Extract the number of billion parameters from the model name diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py index bf230bf794c..edc0635a741 100644 --- a/tests/test_litellm/test_cost_calculator.py +++ b/tests/test_litellm/test_cost_calculator.py @@ -3838,3 +3838,16 @@ def test_completion_cost_together_unmapped_model_still_uses_size_bucket(_local_m ) assert cost == pytest.approx((23 + 15) * 9e-07, rel=1e-9) + + +def test_completion_cost_together_metadata_only_model_still_uses_size_bucket(_local_model_cost_map): + assert "input_cost_per_token" not in litellm.model_cost["together_ai/togethercomputer/CodeLlama-34b-Instruct"] + + cost = completion_cost( + completion_response=_together_chat_response( + model="togethercomputer/CodeLlama-34b-Instruct", prompt_tokens=23, completion_tokens=15, cached_tokens=0 + ), + custom_llm_provider="together_ai", + ) + + assert cost == pytest.approx((23 + 15) * 8e-07, rel=1e-9) From 599905356f44b8f6b12f802e49912f077ceac445 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:17:47 -0700 Subject: [PATCH 14/44] fix(together_ai): pass reasoning_effort=high through on DeepSeek-V4-Pro --- litellm/llms/together_ai/chat/transformation.py | 2 +- .../together_ai/chat/test_together_ai_chat_transformation.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/litellm/llms/together_ai/chat/transformation.py b/litellm/llms/together_ai/chat/transformation.py index cb630711ffe..2f7a7caecd4 100644 --- a/litellm/llms/together_ai/chat/transformation.py +++ b/litellm/llms/together_ai/chat/transformation.py @@ -48,7 +48,7 @@ HYBRID_REASONING_MODELS: Final = frozenset( HIGH_MAX_EFFORT_MODEL_PREFIX: Final = "deepseek-ai/DeepSeek-V4-Pro" EFFORT_TRANSLATION: Final = MappingProxyType({"minimal": "low", "xhigh": "high", "max": "high"}) HIGH_MAX_EFFORT_TRANSLATION: Final = MappingProxyType( - {"minimal": "high", "low": "high", "medium": "high", "high": "max", "xhigh": "max"} + {"minimal": "high", "low": "high", "medium": "high", "xhigh": "max"} ) diff --git a/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py b/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py index d21baa0b43d..f85cb6e05dd 100644 --- a/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py +++ b/tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py @@ -250,7 +250,7 @@ def test_reasoning_effort_none_does_not_clobber_user_reasoning(): @pytest.mark.parametrize( "effort, expected", - [("minimal", "high"), ("low", "high"), ("medium", "high"), ("high", "max"), ("xhigh", "max"), ("max", "max")], + [("minimal", "high"), ("low", "high"), ("medium", "high"), ("high", "high"), ("xhigh", "max"), ("max", "max")], ) def test_deepseek_v4_pro_remaps_to_high_max(effort, expected): mapped = _map_reasoning_effort(HIGH_MAX_REASONING_MODEL, effort) From d266326a42af668fa73c562b44d0322667265fc3 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 26 Aug 2026 16:56:53 +0000 Subject: [PATCH 15/44] fix(model_prices): azure gpt-4.1-nano retirement date, together deprecations, novita gpt-oss-120b vision flag, fireworks deepseek-v4-pro-0813 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 24 +++++- model_prices_and_context_window.json | 24 +++++- .../test_fireworks_serverless_model_costs.py | 86 +++++++++++++++++++ 3 files changed, 128 insertions(+), 6 deletions(-) create mode 100644 tests/test_litellm/test_fireworks_serverless_model_costs.py diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 267b43a2add..00c840c8df1 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -4681,7 +4681,7 @@ "supports_web_search": false }, "azure/gpt-4.1-nano": { - "deprecation_date": "2026-10-14", + "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 2.5e-08, "input_cost_per_token": 1e-07, "input_cost_per_token_batches": 5e-08, @@ -4714,7 +4714,7 @@ "supports_vision": true }, "azure/gpt-4.1-nano-2025-04-14": { - "deprecation_date": "2026-10-14", + "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 2.5e-08, "input_cost_per_token": 1e-07, "input_cost_per_token_batches": 5e-08, @@ -18577,6 +18577,22 @@ "supports_tool_choice": true, "supports_vision": false }, + "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": { + "cache_read_input_token_cost": 4.4e-08, + "input_cost_per_token": 1.32e-06, + "litellm_provider": "fireworks_ai", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.96e-06, + "source": "https://docs.fireworks.ai/serverless/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, "fireworks_ai/accounts/fireworks/models/firefunction-v2": { "input_cost_per_token": 9e-07, "litellm_provider": "fireworks_ai", @@ -38475,6 +38491,7 @@ "supports_tool_choice": true }, "together_ai/google/gemma-3n-E4B-it": { + "deprecation_date": "2026-08-25", "input_cost_per_token": 6e-08, "litellm_provider": "together_ai", "max_input_tokens": 32768, @@ -38510,6 +38527,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-llama/Llama-Guard-4-12B": { + "deprecation_date": "2026-08-25", "input_cost_per_token": 2e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -47311,7 +47329,7 @@ "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_vision": true, + "supports_vision": false, "supports_system_messages": true, "supports_response_schema": true, "supports_reasoning": true diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 267b43a2add..00c840c8df1 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -4681,7 +4681,7 @@ "supports_web_search": false }, "azure/gpt-4.1-nano": { - "deprecation_date": "2026-10-14", + "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 2.5e-08, "input_cost_per_token": 1e-07, "input_cost_per_token_batches": 5e-08, @@ -4714,7 +4714,7 @@ "supports_vision": true }, "azure/gpt-4.1-nano-2025-04-14": { - "deprecation_date": "2026-10-14", + "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 2.5e-08, "input_cost_per_token": 1e-07, "input_cost_per_token_batches": 5e-08, @@ -18577,6 +18577,22 @@ "supports_tool_choice": true, "supports_vision": false }, + "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": { + "cache_read_input_token_cost": 4.4e-08, + "input_cost_per_token": 1.32e-06, + "litellm_provider": "fireworks_ai", + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 3.96e-06, + "source": "https://docs.fireworks.ai/serverless/pricing", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": false + }, "fireworks_ai/accounts/fireworks/models/firefunction-v2": { "input_cost_per_token": 9e-07, "litellm_provider": "fireworks_ai", @@ -38475,6 +38491,7 @@ "supports_tool_choice": true }, "together_ai/google/gemma-3n-E4B-it": { + "deprecation_date": "2026-08-25", "input_cost_per_token": 6e-08, "litellm_provider": "together_ai", "max_input_tokens": 32768, @@ -38510,6 +38527,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-llama/Llama-Guard-4-12B": { + "deprecation_date": "2026-08-25", "input_cost_per_token": 2e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -47311,7 +47329,7 @@ "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_vision": true, + "supports_vision": false, "supports_system_messages": true, "supports_response_schema": true, "supports_reasoning": true diff --git a/tests/test_litellm/test_fireworks_serverless_model_costs.py b/tests/test_litellm/test_fireworks_serverless_model_costs.py new file mode 100644 index 00000000000..0458af0da0e --- /dev/null +++ b/tests/test_litellm/test_fireworks_serverless_model_costs.py @@ -0,0 +1,86 @@ +""" +Validate the Fireworks AI Serverless entry added for #37274 exists in +`model_prices_and_context_window.json` and that the bare Fireworks model ID +resolves through `get_model_info`. + +Pricing as published at https://docs.fireworks.ai/serverless/pricing +(USD per 1M tokens, uncached input / cached input / output): + + accounts/fireworks/models/deepseek-v4-pro-0813 -> $1.32 / $0.044 / $3.96 +""" + +import json +import os + +import pytest + +import litellm +from litellm.utils import get_model_info + + +@pytest.fixture(scope="module", autouse=True) +def _local_model_cost_map(): + """ + Point litellm at the bundled cost map for the duration of this module + only. ``mp.undo()`` restores both the environment variable and + ``litellm.model_cost`` so nothing leaks into later tests. + """ + mp = pytest.MonkeyPatch() + mp.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + mp.setattr(litellm, "model_cost", litellm.get_model_cost_map(url="")) + get_model_info.cache_clear() + yield + mp.undo() + get_model_info.cache_clear() + + +NEW_ENTRIES = { + "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": { + "input_cost_per_token": 1.32e-06, + "cache_read_input_token_cost": 4.4e-08, + "output_cost_per_token": 3.96e-06, + "max_input_tokens": 1048576, + "max_output_tokens": 131072, + }, +} + + +@pytest.fixture(scope="module") +def model_data(): + json_path = os.path.join( + os.path.dirname(__file__), "../../model_prices_and_context_window.json" + ) + with open(json_path) as f: + return json.load(f) + + +def test_fireworks_serverless_entries_exist(model_data): + """The new prefixed entry carries the pricing and metadata from #37274.""" + for key, expected in NEW_ENTRIES.items(): + assert key in model_data, f"{key} is missing from model_prices_and_context_window.json" + entry = model_data[key] + for field, value in expected.items(): + assert entry[field] == pytest.approx(value), f"{key}.{field}" + assert entry["litellm_provider"] == "fireworks_ai" + assert entry["mode"] == "chat" + assert entry["supports_function_calling"] is True + assert entry["supports_vision"] is False + + +def test_bare_fireworks_ids_resolve_through_prefixed_entries(): + """Bare IDs from #37274 resolve via the provider-prefix lookup path.""" + for bare_id, prefixed_key in [ + ( + "accounts/fireworks/models/deepseek-v4-pro-0813", + "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813", + ), + ]: + info = get_model_info(model=bare_id, custom_llm_provider="fireworks_ai") + expected = NEW_ENTRIES[prefixed_key] + assert info.get("key") == prefixed_key + assert info["litellm_provider"] == "fireworks_ai" + assert info["input_cost_per_token"] == pytest.approx(expected["input_cost_per_token"]) + assert info["cache_read_input_token_cost"] == pytest.approx(expected["cache_read_input_token_cost"]) + assert info["output_cost_per_token"] == pytest.approx(expected["output_cost_per_token"]) + assert info["max_input_tokens"] == expected["max_input_tokens"] + assert info["max_output_tokens"] == expected["max_output_tokens"] From d3ede97189e6c65157da5628a195e26972d0ec95 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 26 Aug 2026 16:57:26 +0000 Subject: [PATCH 16/44] fix(model_prices): keep novita gpt-oss-120b vision flag per provider catalog Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/model_prices_and_context_window_backup.json | 2 +- model_prices_and_context_window.json | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 00c840c8df1..1ca0e748d38 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -47329,7 +47329,7 @@ "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_vision": false, + "supports_vision": true, "supports_system_messages": true, "supports_response_schema": true, "supports_reasoning": true diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 00c840c8df1..1ca0e748d38 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -47329,7 +47329,7 @@ "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_tool_choice": true, - "supports_vision": false, + "supports_vision": true, "supports_system_messages": true, "supports_response_schema": true, "supports_reasoning": true From 07c9812739b06d8d972f78fb141fedc834040b57 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 26 Aug 2026 17:18:51 +0000 Subject: [PATCH 17/44] fix(model_prices): carry anthropic behavior flags on deepinfra claude entries, move retired together models to deprecated list Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 84 +++++++++++-------- model_prices_and_context_window.json | 84 +++++++++++-------- .../test_together_ai_model_metadata.py | 4 +- 3 files changed, 102 insertions(+), 70 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 1ca0e748d38..5f565b84d38 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -52689,18 +52689,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-4-8": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 1024, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/anthropic/claude-sonnet-4-6": { "max_tokens": 1000000, @@ -52890,18 +52893,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 512, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/thinkingmachines/Inkling": { "max_tokens": 524288, @@ -53199,18 +53205,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-sonnet-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 2e-06, - "output_cost_per_token": 1e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 1e-05, + "prompt_cache_min_tokens": 1024, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/Qwen/Qwen3.5-397B-A17B": { "max_tokens": 262144, @@ -53288,18 +53297,22 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-fable-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 1e-05, - "output_cost_per_token": 5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 5e-05, + "prompt_cache_min_tokens": 512, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "thinking_always_on": true }, "deepinfra/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { "max_tokens": 262144, @@ -53408,18 +53421,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-4-7": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 2048, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/Qwen/Qwen3.6-27B": { "max_tokens": 262144, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 1ca0e748d38..5f565b84d38 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -52689,18 +52689,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-4-8": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 1024, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/anthropic/claude-sonnet-4-6": { "max_tokens": 1000000, @@ -52890,18 +52893,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 512, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/thinkingmachines/Inkling": { "max_tokens": 524288, @@ -53199,18 +53205,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-sonnet-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 2e-06, - "output_cost_per_token": 1e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 1e-05, + "prompt_cache_min_tokens": 1024, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/Qwen/Qwen3.5-397B-A17B": { "max_tokens": 262144, @@ -53288,18 +53297,22 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-fable-5": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 1e-05, - "output_cost_per_token": 5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 5e-05, + "prompt_cache_min_tokens": 512, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "thinking_always_on": true }, "deepinfra/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": { "max_tokens": 262144, @@ -53408,18 +53421,21 @@ "source": "https://deepinfra.com/pricing" }, "deepinfra/anthropic/claude-opus-4-7": { - "max_tokens": 1000000, - "max_input_tokens": 1000000, "input_cost_per_token": 5e-06, - "output_cost_per_token": 2.5e-05, "litellm_provider": "deepinfra", + "max_input_tokens": 1000000, + "max_tokens": 1000000, "mode": "chat", + "output_cost_per_token": 2.5e-05, + "prompt_cache_min_tokens": 2048, + "source": "https://deepinfra.com/pricing", + "supports_adaptive_thinking": true, "supports_function_calling": true, - "supports_tool_choice": true, - "supports_response_schema": true, "supports_reasoning": true, - "supports_vision": true, - "source": "https://deepinfra.com/pricing" + "supports_response_schema": true, + "supports_sampling_params": false, + "supports_tool_choice": true, + "supports_vision": true }, "deepinfra/Qwen/Qwen3.6-27B": { "max_tokens": 262144, diff --git a/tests/test_litellm/test_together_ai_model_metadata.py b/tests/test_litellm/test_together_ai_model_metadata.py index 5a0aadf4737..60e4b8ddb4d 100644 --- a/tests/test_litellm/test_together_ai_model_metadata.py +++ b/tests/test_litellm/test_together_ai_model_metadata.py @@ -31,16 +31,16 @@ SERVERLESS_CHAT_MODELS: Final = ( "together_ai/meta-models/Muse-Glimmer-30B", "together_ai/google/gemma-4-31B-it", "together_ai/pearl-ai/gemma-4-31b-it", - "together_ai/google/gemma-3n-E4B-it", "together_ai/arize-ai/qwen-2-1.5b-instruct", "together_ai/Prism-ML/Ternary-Bonsai-27B", - "together_ai/meta-llama/Llama-Guard-4-12B", "together_ai/openai/gpt-oss-120b", "together_ai/openai/gpt-oss-20b", "together_ai/meta-llama/Llama-3.3-70B-Instruct-Turbo", ) DEPRECATED_MODELS: Final = { + "together_ai/google/gemma-3n-E4B-it": "2026-08-25", + "together_ai/meta-llama/Llama-Guard-4-12B": "2026-08-25", "together_ai/Qwen/Qwen3-235B-A22B-Instruct-2507-tput": "2026-07-10", "together_ai/Qwen/Qwen3.5-397B-A17B": "2026-06-29", "together_ai/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8": "2026-06-04", From f3c1e2e2a7c053694beb3a48ebb84069120ad239 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 26 Aug 2026 18:12:15 +0000 Subject: [PATCH 18/44] fix(guardrails): forward aws_external_id when the bedrock guardrail assumes a role Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../guardrail_hooks/bedrock_guardrails.py | 2 + .../guardrails/guardrail_initializers.py | 1 + litellm/types/guardrails.py | 3 + .../test_bedrock_guardrails.py | 71 +++++++++++++++++++ 4 files changed, 77 insertions(+) diff --git a/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py b/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py index c70a2ee8a74..dd76a27c80f 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py +++ b/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py @@ -686,6 +686,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM): aws_profile_name: Final = self.optional_params.get("aws_profile_name", None) aws_web_identity_token: Final = self.optional_params.get("aws_web_identity_token", None) aws_sts_endpoint: Final = self.optional_params.get("aws_sts_endpoint", None) + aws_external_id: Final = self.optional_params.get("aws_external_id", None) ### SET REGION NAME ### aws_region_name = self.get_aws_region_name_for_non_llm_api_calls( @@ -702,6 +703,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM): aws_role_name=aws_role_name, aws_web_identity_token=aws_web_identity_token, aws_sts_endpoint=aws_sts_endpoint, + aws_external_id=aws_external_id, ) return credentials, aws_region_name diff --git a/litellm/proxy/guardrails/guardrail_initializers.py b/litellm/proxy/guardrails/guardrail_initializers.py index 0d23e19f88d..b377b272e5b 100644 --- a/litellm/proxy/guardrails/guardrail_initializers.py +++ b/litellm/proxy/guardrails/guardrail_initializers.py @@ -34,6 +34,7 @@ def initialize_bedrock(litellm_params: LitellmParams, guardrail: Guardrail): aws_role_name=litellm_params.aws_role_name, aws_web_identity_token=litellm_params.aws_web_identity_token, aws_sts_endpoint=litellm_params.aws_sts_endpoint, + aws_external_id=litellm_params.aws_external_id, aws_bedrock_runtime_endpoint=litellm_params.aws_bedrock_runtime_endpoint, experimental_use_latest_role_message_only=litellm_params.experimental_use_latest_role_message_only, only_scan_new_messages=litellm_params.only_scan_new_messages or False, diff --git a/litellm/types/guardrails.py b/litellm/types/guardrails.py index c7cdfaad780..5e21dd2f60c 100644 --- a/litellm/types/guardrails.py +++ b/litellm/types/guardrails.py @@ -496,6 +496,9 @@ class BedrockGuardrailConfigModel(BaseModel): aws_role_name: str | None = Field(default=None, description="AWS role name for assuming roles") aws_web_identity_token: str | None = Field(default=None, description="Web identity token for AWS role assumption") aws_sts_endpoint: str | None = Field(default=None, description="AWS STS endpoint URL") + aws_external_id: str | None = Field( + default=None, description="External ID required by the target role's trust policy on sts:AssumeRole" + ) aws_bedrock_runtime_endpoint: str | None = Field(default=None, description="AWS Bedrock runtime endpoint URL") checks: BedrockChecksConfigModel | None = Field( default=None, diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py index dd339d4e51f..36b356e34d0 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py @@ -5274,3 +5274,74 @@ async def test_terminal_failure_logs_usage_and_cost_of_prior_passed_chunks(monke assert logged["guardrail_cost"] == pytest.approx(0.0003) assert logged["guardrail_response"]["usage"] == {"contentPolicyUnits": 2, "wordPolicyUnits": 1} assert "error" in logged["guardrail_response"] + + +def test_load_credentials_assumes_role_with_external_id(): + """A trust policy requiring sts:ExternalId must be satisfied by the guardrail's aws_external_id.""" + import datetime + + import boto3 + from botocore.exceptions import ClientError + + class FakeSTSClient: + """STS that mirrors a cross-account role whose trust policy requires an ExternalId.""" + + def get_caller_identity(self): + return {"Arn": "arn:aws:iam::111111111111:user/litellm-proxy-pod"} + + def assume_role(self, **params): + if params.get("ExternalId") != "external-id-123": + raise ClientError( + {"Error": {"Code": "AccessDenied", "Message": "is not authorized to perform: sts:AssumeRole"}}, + "AssumeRole", + ) + return { + "Credentials": { + "AccessKeyId": "ASIAASSUMEDROLEKEY", + "SecretAccessKey": "assumed-secret", + "SessionToken": "assumed-session-token", + "Expiration": datetime.datetime.now(datetime.timezone.utc) + datetime.timedelta(minutes=30), + } + } + + guardrail = BedrockGuardrail( + guardrail_name="bedrock-external-id", + event_hook=GuardrailEventHooks.pre_call, + guardrailIdentifier="gr-1", + guardrailVersion="DRAFT", + aws_region_name="us-east-1", + aws_access_key_id="AKIAPODCALLERKEY", + aws_secret_access_key="pod-caller-secret", + aws_role_name="arn:aws:iam::999999999999:role/litellm-guardrail-role", + aws_session_name="litellm-session", + aws_external_id="external-id-123", + ) + + with patch.object(boto3, "client", return_value=FakeSTSClient()): + credentials, aws_region_name = guardrail._load_credentials() + + assert credentials.access_key == "ASIAASSUMEDROLEKEY" + assert credentials.token == "assumed-session-token" + assert aws_region_name == "us-east-1" + + +def test_initialize_bedrock_forwards_aws_external_id(): + """aws_external_id configured on the guardrail must survive LitellmParams and the initializer.""" + from litellm.proxy.guardrails.guardrail_initializers import initialize_bedrock + from litellm.types.guardrails import LitellmParams + + litellm_params = LitellmParams( + guardrail="bedrock", + mode="pre_call", + guardrailIdentifier="gr-1", + guardrailVersion="DRAFT", + aws_region_name="us-east-1", + aws_role_name="arn:aws:iam::999999999999:role/litellm-guardrail-role", + aws_external_id="external-id-123", + ) + + guardrail = initialize_bedrock(litellm_params, {"guardrail_name": "bedrock-external-id"}) + try: + assert guardrail.optional_params["aws_external_id"] == "external-id-123" + finally: + litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, guardrail) From 4456a4407feb474439d344ee5116d7d82d8701ab Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Wed, 26 Aug 2026 19:20:14 +0000 Subject: [PATCH 19/44] fix(model_prices): azure gpt-5.6 cache writes, mistral missing models, together cache reads Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 272 ++++++++++++++++-- model_prices_and_context_window.json | 272 ++++++++++++++++-- tests/test_litellm/test_cost_calculator.py | 78 +++++ 3 files changed, 560 insertions(+), 62 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 7515e7ad396..1526319ecf1 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -6592,6 +6592,9 @@ "supports_web_search": true }, "azure/gpt-5.6": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, + "cache_creation_input_token_cost_priority": 1.25e-05, "cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_above_272k_tokens": 1e-06, "cache_read_input_token_cost_priority": 1e-06, @@ -6642,6 +6645,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, + "cache_creation_input_token_cost_priority": 1.25e-05, "cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_above_272k_tokens": 1e-06, "cache_read_input_token_cost_priority": 1e-06, @@ -6693,6 +6699,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5e-06, + "cache_creation_input_token_cost_priority": 5e-06, "cache_read_input_token_cost": 2e-07, "cache_read_input_token_cost_above_272k_tokens": 4e-07, "cache_read_input_token_cost_priority": 4e-07, @@ -6744,6 +6753,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.5e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5e-07, + "cache_creation_input_token_cost_priority": 5e-07, "cache_read_input_token_cost": 2e-08, "cache_read_input_token_cost_above_272k_tokens": 4e-08, "cache_read_input_token_cost_priority": 4e-08, @@ -6795,12 +6807,15 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6808,7 +6823,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6842,13 +6857,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "deprecation_date": "2028-01-11", "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6856,7 +6874,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6890,13 +6908,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.75e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06, + "cache_creation_input_token_cost_priority": 5.5e-06, "cache_read_input_token_cost": 2.2e-07, "cache_read_input_token_cost_above_272k_tokens": 4.4e-07, - "cache_read_input_token_cost_priority": 5.5e-07, + "cache_read_input_token_cost_priority": 4.4e-07, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-06, "input_cost_per_token_above_272k_tokens": 4.4e-06, - "input_cost_per_token_priority": 5.5e-06, + "input_cost_per_token_priority": 4.4e-06, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6904,7 +6925,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-05, "output_cost_per_token_above_272k_tokens": 1.98e-05, - "output_cost_per_token_priority": 3.3e-05, + "output_cost_per_token_priority": 2.64e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6938,13 +6959,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.75e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07, + "cache_creation_input_token_cost_priority": 5.5e-07, "cache_read_input_token_cost": 2.2e-08, "cache_read_input_token_cost_above_272k_tokens": 4.4e-08, - "cache_read_input_token_cost_priority": 5.5e-08, + "cache_read_input_token_cost_priority": 4.4e-08, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, - "input_cost_per_token_priority": 5.5e-07, + "input_cost_per_token_priority": 4.4e-07, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6952,7 +6976,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-06, "output_cost_per_token_above_272k_tokens": 1.98e-06, - "output_cost_per_token_priority": 3.3e-06, + "output_cost_per_token_priority": 2.64e-06, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6986,12 +7010,15 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6999,7 +7026,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7033,13 +7060,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "deprecation_date": "2028-01-11", "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7047,7 +7077,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7081,13 +7111,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.75e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06, + "cache_creation_input_token_cost_priority": 5.5e-06, "cache_read_input_token_cost": 2.2e-07, "cache_read_input_token_cost_above_272k_tokens": 4.4e-07, - "cache_read_input_token_cost_priority": 5.5e-07, + "cache_read_input_token_cost_priority": 4.4e-07, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-06, "input_cost_per_token_above_272k_tokens": 4.4e-06, - "input_cost_per_token_priority": 5.5e-06, + "input_cost_per_token_priority": 4.4e-06, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7095,7 +7128,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-05, "output_cost_per_token_above_272k_tokens": 1.98e-05, - "output_cost_per_token_priority": 3.3e-05, + "output_cost_per_token_priority": 2.64e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7129,13 +7162,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.75e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07, + "cache_creation_input_token_cost_priority": 5.5e-07, "cache_read_input_token_cost": 2.2e-08, "cache_read_input_token_cost_above_272k_tokens": 4.4e-08, - "cache_read_input_token_cost_priority": 5.5e-08, + "cache_read_input_token_cost_priority": 4.4e-08, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, - "input_cost_per_token_priority": 5.5e-07, + "input_cost_per_token_priority": 4.4e-07, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7143,7 +7179,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-06, "output_cost_per_token_above_272k_tokens": 1.98e-06, - "output_cost_per_token_priority": 3.3e-06, + "output_cost_per_token_priority": 2.64e-06, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -30721,6 +30757,152 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "mistral/ministral-14b-2512": { + "input_cost_per_token": 2e-07, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://docs.mistral.ai/models/ministral-3-14b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-14b-latest": { + "input_cost_per_token": 2e-07, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://docs.mistral.ai/models/ministral-3-14b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-3b-2512": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1e-07, + "source": "https://docs.mistral.ai/models/ministral-3-3b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-3b-latest": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1e-07, + "source": "https://docs.mistral.ai/models/ministral-3-3b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/mistral-embed-2312": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 8192, + "max_tokens": 8192, + "mode": "embedding", + "source": "https://docs.mistral.ai/models/mistral-embed-23-12" + }, + "mistral/mistral-medium-3": { + "input_cost_per_token": 1.5e-06, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 7.5e-06, + "source": "https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/voxtral-mini-transcribe-realtime-latest": { + "input_cost_per_second": 0.0001, + "litellm_provider": "mistral", + "mode": "audio_transcription", + "source": "https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-26-02", + "supported_endpoints": [ + "/v1/audio/transcriptions" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "mistral/voxtral-mini-tts-latest": { + "litellm_provider": "mistral", + "mode": "audio_speech", + "output_cost_per_character": 1.6e-05, + "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03", + "supported_endpoints": [ + "/v1/audio/speech" + ], + "supported_modalities": [ + "text" + ], + "supported_output_modalities": [ + "audio" + ], + "supports_audio_output": true + }, + "mistral/voxtral-small-2507": { + "input_cost_per_second": 6.666666666666667e-05, + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 32768, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 4e-07, + "source": "https://docs.mistral.ai/models/voxtral-small-25-07", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "mistral/voxtral-small-latest": { + "input_cost_per_second": 6.666666666666667e-05, + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 32768, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 4e-07, + "source": "https://docs.mistral.ai/models/voxtral-small-25-07", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, "mistral/zai-glm-5-2": { "cache_read_input_token_cost": 1.4e-07, "input_cost_per_token": 1.4e-06, @@ -31191,9 +31373,9 @@ "mistral/ministral-3-3b-2512": { "input_cost_per_token": 1e-07, "litellm_provider": "mistral", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 1e-07, "source": "https://mistral.ai/pricing", @@ -38342,6 +38524,7 @@ "supports_tool_choice": true }, "together_ai/Qwen/Qwen3.5-397B-A17B": { + "cache_read_input_token_cost": 3.5e-07, "deprecation_date": "2026-06-29", "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", @@ -38351,10 +38534,12 @@ "source": "https://www.together.ai/models/Qwen/Qwen3.5-397B-A17B", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/MiniMaxAI/MiniMax-M3": { + "cache_read_input_token_cost": 6e-08, "input_cost_per_token": 3e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38365,6 +38550,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -38408,6 +38594,7 @@ "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.25e-06, "litellm_provider": "together_ai", "max_input_tokens": 1000000, @@ -38415,7 +38602,8 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 3.75e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { "input_cost_per_token": 3.2e-07, @@ -38428,6 +38616,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { + "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1010000, @@ -38435,7 +38624,8 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6.25e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { "input_cost_per_token": 1e-07, @@ -38448,6 +38638,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { + "cache_read_input_token_cost": 3e-08, "input_cost_per_token": 1.4e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38458,10 +38649,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", "max_input_tokens": 512000, @@ -38472,11 +38665,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.32e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38487,6 +38682,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, @@ -38538,6 +38734,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-models/Muse-Glimmer-30B": { + "cache_read_input_token_cost": 4e-08, "input_cost_per_token": 3.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 131072, @@ -38545,9 +38742,11 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -38558,11 +38757,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/moonshotai/Kimi-K3": { + "cache_read_input_token_cost": 3e-07, "input_cost_per_token": 3e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38573,12 +38774,14 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", "max_input_tokens": 512288, @@ -38589,6 +38792,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true @@ -38604,6 +38808,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/thinkingmachines/Inkling": { + "cache_read_input_token_cost": 1.7e-07, "input_cost_per_token": 1e-06, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38614,10 +38819,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/thinkingmachines/Inkling-Small": { + "cache_read_input_token_cost": 1e-07, "input_cost_per_token": 5e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38625,9 +38832,11 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { + "cache_read_input_token_cost": 2.6e-07, "input_cost_per_token": 1.4e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048575, @@ -38638,6 +38847,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 7515e7ad396..1526319ecf1 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -6592,6 +6592,9 @@ "supports_web_search": true }, "azure/gpt-5.6": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, + "cache_creation_input_token_cost_priority": 1.25e-05, "cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_above_272k_tokens": 1e-06, "cache_read_input_token_cost_priority": 1e-06, @@ -6642,6 +6645,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.25e-05, + "cache_creation_input_token_cost_priority": 1.25e-05, "cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_above_272k_tokens": 1e-06, "cache_read_input_token_cost_priority": 1e-06, @@ -6693,6 +6699,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.5e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5e-06, + "cache_creation_input_token_cost_priority": 5e-06, "cache_read_input_token_cost": 2e-07, "cache_read_input_token_cost_above_272k_tokens": 4e-07, "cache_read_input_token_cost_priority": 4e-07, @@ -6744,6 +6753,9 @@ "supports_minimal_reasoning_effort": false }, "azure/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.5e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5e-07, + "cache_creation_input_token_cost_priority": 5e-07, "cache_read_input_token_cost": 2e-08, "cache_read_input_token_cost_above_272k_tokens": 4e-08, "cache_read_input_token_cost_priority": 4e-08, @@ -6795,12 +6807,15 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6808,7 +6823,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6842,13 +6857,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "deprecation_date": "2028-01-11", "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6856,7 +6874,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6890,13 +6908,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.75e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06, + "cache_creation_input_token_cost_priority": 5.5e-06, "cache_read_input_token_cost": 2.2e-07, "cache_read_input_token_cost_above_272k_tokens": 4.4e-07, - "cache_read_input_token_cost_priority": 5.5e-07, + "cache_read_input_token_cost_priority": 4.4e-07, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-06, "input_cost_per_token_above_272k_tokens": 4.4e-06, - "input_cost_per_token_priority": 5.5e-06, + "input_cost_per_token_priority": 4.4e-06, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6904,7 +6925,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-05, "output_cost_per_token_above_272k_tokens": 1.98e-05, - "output_cost_per_token_priority": 3.3e-05, + "output_cost_per_token_priority": 2.64e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6938,13 +6959,16 @@ "supports_minimal_reasoning_effort": false }, "azure/us/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.75e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07, + "cache_creation_input_token_cost_priority": 5.5e-07, "cache_read_input_token_cost": 2.2e-08, "cache_read_input_token_cost_above_272k_tokens": 4.4e-08, - "cache_read_input_token_cost_priority": 5.5e-08, + "cache_read_input_token_cost_priority": 4.4e-08, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, - "input_cost_per_token_priority": 5.5e-07, + "input_cost_per_token_priority": 4.4e-07, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6952,7 +6976,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-06, "output_cost_per_token_above_272k_tokens": 1.98e-06, - "output_cost_per_token_priority": 3.3e-06, + "output_cost_per_token_priority": 2.64e-06, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -6986,12 +7010,15 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -6999,7 +7026,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7033,13 +7060,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-sol": { + "cache_creation_input_token_cost": 6.875e-06, + "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05, + "cache_creation_input_token_cost_priority": 1.375e-05, "cache_read_input_token_cost": 5.5e-07, "cache_read_input_token_cost_above_272k_tokens": 1.1e-06, - "cache_read_input_token_cost_priority": 1.375e-06, + "cache_read_input_token_cost_priority": 1.1e-06, "deprecation_date": "2028-01-11", "input_cost_per_token": 5.5e-06, "input_cost_per_token_above_272k_tokens": 1.1e-05, - "input_cost_per_token_priority": 1.375e-05, + "input_cost_per_token_priority": 1.1e-05, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7047,7 +7077,7 @@ "mode": "chat", "output_cost_per_token": 3.3e-05, "output_cost_per_token_above_272k_tokens": 4.95e-05, - "output_cost_per_token_priority": 8.25e-05, + "output_cost_per_token_priority": 6.6e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7081,13 +7111,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-terra": { + "cache_creation_input_token_cost": 2.75e-06, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06, + "cache_creation_input_token_cost_priority": 5.5e-06, "cache_read_input_token_cost": 2.2e-07, "cache_read_input_token_cost_above_272k_tokens": 4.4e-07, - "cache_read_input_token_cost_priority": 5.5e-07, + "cache_read_input_token_cost_priority": 4.4e-07, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-06, "input_cost_per_token_above_272k_tokens": 4.4e-06, - "input_cost_per_token_priority": 5.5e-06, + "input_cost_per_token_priority": 4.4e-06, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7095,7 +7128,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-05, "output_cost_per_token_above_272k_tokens": 1.98e-05, - "output_cost_per_token_priority": 3.3e-05, + "output_cost_per_token_priority": 2.64e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -7129,13 +7162,16 @@ "supports_minimal_reasoning_effort": false }, "azure/eu/gpt-5.6-luna": { + "cache_creation_input_token_cost": 2.75e-07, + "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07, + "cache_creation_input_token_cost_priority": 5.5e-07, "cache_read_input_token_cost": 2.2e-08, "cache_read_input_token_cost_above_272k_tokens": 4.4e-08, - "cache_read_input_token_cost_priority": 5.5e-08, + "cache_read_input_token_cost_priority": 4.4e-08, "deprecation_date": "2028-01-11", "input_cost_per_token": 2.2e-07, "input_cost_per_token_above_272k_tokens": 4.4e-07, - "input_cost_per_token_priority": 5.5e-07, + "input_cost_per_token_priority": 4.4e-07, "litellm_provider": "azure", "max_input_tokens": 922000, "max_output_tokens": 128000, @@ -7143,7 +7179,7 @@ "mode": "chat", "output_cost_per_token": 1.32e-06, "output_cost_per_token_above_272k_tokens": 1.98e-06, - "output_cost_per_token_priority": 3.3e-06, + "output_cost_per_token_priority": 2.64e-06, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -30721,6 +30757,152 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "mistral/ministral-14b-2512": { + "input_cost_per_token": 2e-07, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://docs.mistral.ai/models/ministral-3-14b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-14b-latest": { + "input_cost_per_token": 2e-07, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://docs.mistral.ai/models/ministral-3-14b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-3b-2512": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1e-07, + "source": "https://docs.mistral.ai/models/ministral-3-3b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/ministral-3b-latest": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, + "mode": "chat", + "output_cost_per_token": 1e-07, + "source": "https://docs.mistral.ai/models/ministral-3-3b-25-12", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/mistral-embed-2312": { + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 8192, + "max_tokens": 8192, + "mode": "embedding", + "source": "https://docs.mistral.ai/models/mistral-embed-23-12" + }, + "mistral/mistral-medium-3": { + "input_cost_per_token": 1.5e-06, + "litellm_provider": "mistral", + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, + "mode": "chat", + "output_cost_per_token": 7.5e-06, + "source": "https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04", + "supports_assistant_prefill": true, + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "mistral/voxtral-mini-transcribe-realtime-latest": { + "input_cost_per_second": 0.0001, + "litellm_provider": "mistral", + "mode": "audio_transcription", + "source": "https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-26-02", + "supported_endpoints": [ + "/v1/audio/transcriptions" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "mistral/voxtral-mini-tts-latest": { + "litellm_provider": "mistral", + "mode": "audio_speech", + "output_cost_per_character": 1.6e-05, + "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03", + "supported_endpoints": [ + "/v1/audio/speech" + ], + "supported_modalities": [ + "text" + ], + "supported_output_modalities": [ + "audio" + ], + "supports_audio_output": true + }, + "mistral/voxtral-small-2507": { + "input_cost_per_second": 6.666666666666667e-05, + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 32768, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 4e-07, + "source": "https://docs.mistral.ai/models/voxtral-small-25-07", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "mistral/voxtral-small-latest": { + "input_cost_per_second": 6.666666666666667e-05, + "input_cost_per_token": 1e-07, + "litellm_provider": "mistral", + "max_input_tokens": 32768, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 4e-07, + "source": "https://docs.mistral.ai/models/voxtral-small-25-07", + "supports_audio_input": true, + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, "mistral/zai-glm-5-2": { "cache_read_input_token_cost": 1.4e-07, "input_cost_per_token": 1.4e-06, @@ -31191,9 +31373,9 @@ "mistral/ministral-3-3b-2512": { "input_cost_per_token": 1e-07, "litellm_provider": "mistral", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, + "max_input_tokens": 262144, + "max_output_tokens": 262144, + "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 1e-07, "source": "https://mistral.ai/pricing", @@ -38342,6 +38524,7 @@ "supports_tool_choice": true }, "together_ai/Qwen/Qwen3.5-397B-A17B": { + "cache_read_input_token_cost": 3.5e-07, "deprecation_date": "2026-06-29", "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", @@ -38351,10 +38534,12 @@ "source": "https://www.together.ai/models/Qwen/Qwen3.5-397B-A17B", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/MiniMaxAI/MiniMax-M3": { + "cache_read_input_token_cost": 6e-08, "input_cost_per_token": 3e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38365,6 +38550,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -38408,6 +38594,7 @@ "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.25e-06, "litellm_provider": "together_ai", "max_input_tokens": 1000000, @@ -38415,7 +38602,8 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 3.75e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { "input_cost_per_token": 3.2e-07, @@ -38428,6 +38616,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { + "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2.5e-06, "litellm_provider": "together_ai", "max_input_tokens": 1010000, @@ -38435,7 +38624,8 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6.25e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { "input_cost_per_token": 1e-07, @@ -38448,6 +38638,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { + "cache_read_input_token_cost": 3e-08, "input_cost_per_token": 1.4e-07, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38458,10 +38649,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", "max_input_tokens": 512000, @@ -38472,11 +38665,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813": { + "cache_read_input_token_cost": 1.3e-07, "input_cost_per_token": 1.32e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38487,6 +38682,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, @@ -38538,6 +38734,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/meta-models/Muse-Glimmer-30B": { + "cache_read_input_token_cost": 4e-08, "input_cost_per_token": 3.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 131072, @@ -38545,9 +38742,11 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -38558,11 +38757,13 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/moonshotai/Kimi-K3": { + "cache_read_input_token_cost": 3e-07, "input_cost_per_token": 3e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048576, @@ -38573,12 +38774,14 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", "max_input_tokens": 512288, @@ -38589,6 +38792,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true @@ -38604,6 +38808,7 @@ "source": "https://docs.together.ai/docs/serverless-models" }, "together_ai/thinkingmachines/Inkling": { + "cache_read_input_token_cost": 1.7e-07, "input_cost_per_token": 1e-06, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38614,10 +38819,12 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true }, "together_ai/thinkingmachines/Inkling-Small": { + "cache_read_input_token_cost": 1e-07, "input_cost_per_token": 5e-07, "litellm_provider": "together_ai", "max_input_tokens": 524288, @@ -38625,9 +38832,11 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://docs.together.ai/docs/serverless-models", + "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { + "cache_read_input_token_cost": 2.6e-07, "input_cost_per_token": 1.4e-06, "litellm_provider": "together_ai", "max_input_tokens": 1048575, @@ -38638,6 +38847,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py index 3d0921c6fd8..578282efee5 100644 --- a/tests/test_litellm/test_cost_calculator.py +++ b/tests/test_litellm/test_cost_calculator.py @@ -1727,6 +1727,84 @@ def test_azure_ai_cache_cost_calculation(_local_model_cost_map): ), f"Output cost mismatch: got {output_cost}, expected {expected_output_cost}" +AZURE_GPT_5_6_MAP_KEYS = ( + "azure/gpt-5.6", + "azure/gpt-5.6-sol", + "azure/gpt-5.6-terra", + "azure/gpt-5.6-luna", + "azure/us/gpt-5.6", + "azure/us/gpt-5.6-sol", + "azure/us/gpt-5.6-terra", + "azure/us/gpt-5.6-luna", + "azure/eu/gpt-5.6", + "azure/eu/gpt-5.6-sol", + "azure/eu/gpt-5.6-terra", + "azure/eu/gpt-5.6-luna", +) + + +def test_azure_gpt_5_6_cache_write_tokens_are_billed(_local_model_cost_map): + """ + Azure bills gpt-5.6 prompt cache writes at 1.25x the input rate, but the + azure entries carried no ``cache_creation_input_token_cost``, so + cache-write tokens were billed at the plain input rate instead. + """ + from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token + from litellm.types.utils import PromptTokensDetailsWrapper, Usage + + usage = Usage( + completion_tokens=100, + prompt_tokens=2000, + total_tokens=2100, + prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=0, text_tokens=687), + cache_creation_input_tokens=1313, + ) + + input_cost, output_cost = generic_cost_per_token( + model="azure/gpt-5.6-luna", usage=usage, custom_llm_provider="azure" + ) + + assert input_cost == pytest.approx(687 * 2e-07 + 1313 * 2.5e-07) + assert output_cost == pytest.approx(100 * 1.2e-06) + + +@pytest.mark.parametrize("model", AZURE_GPT_5_6_MAP_KEYS) +def test_azure_gpt_5_6_rates_match_azure_price_page(_local_model_cost_map, model): + """ + Per the Azure retail price API (2026-08-26): cache writes cost 1.25x input + on every published gpt-5.6 meter, and Data Zone standard and priority + rates cost 1.1x Global (us/eu priority rates previously sat at 1.25x). + Azure publishes no long-context priority meters, so the + ``*_above_272k_tokens_priority`` suffix is excluded. + """ + entry = litellm.model_cost[model] + input_keys = [ + key + for key in entry + if key.startswith("input_cost_per_token") + and not key.endswith("_above_272k_tokens_priority") + ] + assert input_keys + for key in input_keys: + suffix = key[len("input_cost_per_token") :] + assert entry["cache_creation_input_token_cost" + suffix] == pytest.approx( + entry[key] * 1.25 + ), key + + zone = model.split("/")[1] + if zone in ("us", "eu"): + global_entry = litellm.model_cost["azure/" + model.split("/", 2)[2]] + prefixes = ("input_cost_per_token", "output_cost_per_token", "cache_read", "cache_creation") + token_cost_keys = [ + key + for key in entry + if key.startswith(prefixes) and not key.endswith("_above_272k_tokens_priority") + ] + assert len(token_cost_keys) >= 9 + for key in token_cost_keys: + assert entry[key] == pytest.approx(global_entry[key] * 1.1), key + + def test_vertex_regional_deployment_costs_uplift_over_global(monkeypatch): """ Regression for https://github.com/BerriAI/litellm/issues/34393: two Vertex From f7af44a5055e55bb6c7d3be0a6154daaad14fe4b Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 17:55:15 -0700 Subject: [PATCH 20/44] fix(anthropic-adapter): pass provider-native and OpenAI-format tools through on /v1/messages --- .../adapters/transformation.py | 20 ++++++++ ...al_pass_through_adapters_transformation.py | 48 +++++++++++++++++++ 2 files changed, 68 insertions(+) diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index 109017bda27..ca90df2ff75 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -18,6 +18,22 @@ TOOL_NAME_PREFIX_LENGTH: Final = OPENAI_MAX_TOOL_NAME_LENGTH - TOOL_NAME_HASH_LE PROVIDERS_PROXYING_AN_UNKNOWN_BACKEND: Final = frozenset({"litellm_proxy"}) +_ANTHROPIC_TOOL_SCHEMA_KEYS: Final = frozenset( + {"name", "type", "input_schema", "description", "cache_control", "strict"} +) + + +def _is_openai_function_tool(tool: Mapping[str, object]) -> bool: + return tool.get("type") == "function" and "function" in tool + + +def _is_provider_native_tool_dict(tool: Mapping[str, object]) -> bool: + if len(tool) != 1: + return False + key, value = next(iter(tool.items())) + return key not in _ANTHROPIC_TOOL_SCHEMA_KEYS and isinstance(value, dict) + + def truncate_tool_name(name: str) -> str: """ Truncate tool names that exceed OpenAI's 64-character limit. @@ -770,6 +786,10 @@ class LiteLLMAnthropicMessagesAdapter: new_tools.append(tool) continue + if _is_openai_function_tool(tool) or _is_provider_native_tool_dict(tool): + new_tools.append(cast(ChatCompletionToolParam, tool)) # cast-ok: passed through verbatim to provider + continue + raw_name = tool.get("name") if raw_name is None or (isinstance(raw_name, str) and not str(raw_name).strip()): original_name = f"litellm_unnamed_tool_{idx}" diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py index ee09baf28b6..7073d702db4 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py @@ -16,6 +16,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import ( ) from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import ( OPENAI_MAX_TOOL_NAME_LENGTH, + AnthropicAdapter, LiteLLMAnthropicMessagesAdapter, create_tool_name_mapping, truncate_tool_name, @@ -2307,6 +2308,53 @@ def test_translate_anthropic_tools_to_openai_fills_missing_tool_name(): assert result[1]["function"]["name"] == "litellm_unnamed_tool_1" +def test_translate_anthropic_tools_to_openai_passes_provider_native_tool_dicts_through(): + """Deployment-level provider-native tools (e.g. Gemini googleMaps) must reach the provider transformation verbatim (LIT-6286).""" + tools = [ + {"googleMaps": {}}, + {"googleSearch": {}}, + { + "name": "get_weather", + "input_schema": {"type": "object", "properties": {"location": {"type": "string"}}}, + }, + ] + adapter = LiteLLMAnthropicMessagesAdapter() + result, tool_name_mapping = adapter.translate_anthropic_tools_to_openai(tools=tools, model=None) + assert result[0] == {"googleMaps": {}} + assert result[1] == {"googleSearch": {}} + assert result[2]["function"]["name"] == "get_weather" + assert tool_name_mapping == {} + + +def test_translate_anthropic_tools_to_openai_passes_openai_function_tools_through(): + """A tool already in OpenAI function format must pass through unchanged instead of becoming litellm_unnamed_tool_N.""" + openai_tool = { + "type": "function", + "function": { + "name": "get_weather", + "parameters": {"type": "object", "properties": {"location": {"type": "string"}}}, + }, + } + adapter = LiteLLMAnthropicMessagesAdapter() + result, _ = adapter.translate_anthropic_tools_to_openai(tools=[openai_tool], model=None) + assert result == [openai_tool] + + +def test_translate_completion_input_params_keeps_provider_native_tools(): + """/v1/messages request translation must keep router-merged provider-native tools in kwargs['tools'] (LIT-6286).""" + adapter = AnthropicAdapter() + translated = adapter.translate_completion_input_params( + { + "model": "gemini/gemini-2.5-flash", + "max_tokens": 1024, + "messages": [{"role": "user", "content": "coffee shops near Union Square"}], + "tools": [{"googleMaps": {}}], + } + ) + assert translated is not None + assert translated["tools"] == [{"googleMaps": {}}] + + def test_translate_openai_content_to_anthropic_reasoning_content_without_thinking_blocks(): """ Test that reasoning_content is converted to thinking block when thinking_blocks is not present. From ba7c268a6ab496c93b7e0ebd5177bd268f54a30e Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 18:16:12 -0700 Subject: [PATCH 21/44] fix(proxy): enforce tool allowlist on OpenAI-format tools sent to /v1/messages --- .../chat/guardrail_translation/handler.py | 9 +++---- .../base_llm/guardrail_translation/utils.py | 12 +++++++-- .../proxy/test_tools_allowlist_enforcement.py | 27 +++++++++++++++++++ 3 files changed, 40 insertions(+), 8 deletions(-) diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py index 721a6653597..2afea1444f9 100644 --- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py +++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py @@ -677,12 +677,9 @@ class AnthropicMessagesHandler(BaseTranslation): ) def extract_request_tool_names(self, data: dict) -> list[str]: - """Extract tool names from Anthropic messages request (tools[].name).""" - names: Final[list[str]] = [] - for tool in data.get("tools") or []: - if isinstance(tool, dict) and tool.get("name"): - names.append(str(tool["name"])) - return names + """Extract tool names from Anthropic messages request (tools[].name, or + tools[].function.name for OpenAI-format tools the bridge forwards verbatim).""" + return [name for tool in data.get("tools") or [] if (name := anthropic_tool_name(tool))] @classmethod def _extract_input_text_and_images( diff --git a/litellm/llms/base_llm/guardrail_translation/utils.py b/litellm/llms/base_llm/guardrail_translation/utils.py index 1546adbb0bd..4494f8451f4 100644 --- a/litellm/llms/base_llm/guardrail_translation/utils.py +++ b/litellm/llms/base_llm/guardrail_translation/utils.py @@ -210,8 +210,16 @@ def openai_tool_name(tool: object) -> str | None: def anthropic_tool_name(tool: object) -> str | None: - name: Final = tool.get("name") if isinstance(tool, dict) else None - return name if isinstance(name, str) else None + """Anthropic tools carry a flat ``name``; an OpenAI-format function tool, which the + non-Anthropic bridge forwards verbatim, carries it under ``function.name`` instead.""" + if not isinstance(tool, dict): + return None + flat_name: Final = tool.get("name") + if isinstance(flat_name, str): + return flat_name + function: Final = tool.get("function") if tool.get("type") == "function" else None + function_name: Final = function.get("name") if isinstance(function, dict) else None + return function_name if isinstance(function_name, str) else None def merge_returned_tools_into_request_tools( diff --git a/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py b/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py index 31f5fbf606b..c881effada5 100644 --- a/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py +++ b/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py @@ -93,6 +93,19 @@ class TestExtractRequestToolNames: "run_sql", ] + def test_anthropic_openai_format_tools_forwarded_by_bridge(self): + data = { + "tools": [ + {"type": "function", "function": {"name": "get_weather"}}, + {"name": "run_sql"}, + {"googleSearch": {}}, + ] + } + assert extract_request_tool_names("/v1/messages", data) == [ + "get_weather", + "run_sql", + ] + def test_generate_content_tools(self): data = { "tools": [ @@ -159,6 +172,20 @@ class TestCheckToolsAllowlist: assert exc_info.value.type == ProxyErrorTypes.tool_access_denied assert "get_weather" in str(exc_info.value.message) + @pytest.mark.asyncio + async def test_disallowed_openai_format_tool_raises_on_messages_route(self): + token = _token(metadata={"allowed_tools": ["other_tool"]}) + body = {"tools": [{"type": "function", "function": {"name": "get_weather"}}]} + with pytest.raises(ProxyException) as exc_info: + await check_tools_allowlist( + request_body=body, + valid_token=token, + team_object=None, + route="/v1/messages", + ) + assert exc_info.value.type == ProxyErrorTypes.tool_access_denied + assert "get_weather" in str(exc_info.value.message) + @pytest.mark.asyncio async def test_disallowed_custom_tool_raises_on_responses_route(self): token = _token(metadata={"allowed_tools": ["other_tool"]}) From f7c9c87280c55de78abcba429d5cca64342b2890 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 27 Aug 2026 01:59:48 +0000 Subject: [PATCH 22/44] feat(dashscope): support qwen-image-3.0 and qwen-image-3.0-pro image generation Register both models, route image requests to the multimodal generation endpoint instead of the chat compatible-mode base, and pass OpenAI n through as DashScope n so multi-image requests return every image. --- .../image_generation/transformation.py | 19 ++-- ...odel_prices_and_context_window_backup.json | 16 +++ model_prices_and_context_window.json | 16 +++ .../test_dashscope_image_generation.py | 102 ++++++++++++++++-- 4 files changed, 141 insertions(+), 12 deletions(-) diff --git a/litellm/llms/dashscope/image_generation/transformation.py b/litellm/llms/dashscope/image_generation/transformation.py index 9652a5738c8..1347fee897d 100644 --- a/litellm/llms/dashscope/image_generation/transformation.py +++ b/litellm/llms/dashscope/image_generation/transformation.py @@ -11,7 +11,7 @@ Request format: "input": { "messages": [{"role": "user", "content": [{"text": ""}]}] }, - "parameters": {"size": "1024*1024", ...} + "parameters": {"size": "1024*1024", "n": 1, ...} } Response format: @@ -19,7 +19,7 @@ Response format: "output": { "choices": [{"message": {"content": [{"image": ""}]}}] }, - "usage": {"input_tokens": 0, "output_tokens": 0, "width": 1024, "height": 1024, "image_count": 1} + "usage": {"output_width": 1024, "output_height": 1024, "output_image_count": 1} } """ @@ -46,6 +46,9 @@ else: DEFAULT_API_BASE: Final = "https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation" +# get_llm_provider resolves every dashscope route to the chat/embed base, which cannot serve images +CHAT_COMPATIBLE_MODE_PATH: Final = "/compatible-mode/v1" + # Maps OpenAI size strings (WxH) to DashScope size strings (W*H) OPENAI_TO_DASHSCOPE_SIZE: Final[dict] = { "256x256": "256*256", @@ -59,7 +62,8 @@ OPENAI_TO_DASHSCOPE_SIZE: Final[dict] = { class DashScopeImageGenerationConfig(BaseImageGenerationConfig): """ - Configuration for DashScope image generation (qwen-image-2.0, qwen-image-2.0-pro). + Configuration for DashScope image generation (qwen-image-2.0, qwen-image-2.0-pro, + qwen-image-3.0, qwen-image-3.0-pro). """ def get_supported_openai_params(self, model: str) -> list[OpenAIImageGenerationOptionalParams]: @@ -82,8 +86,8 @@ class DashScopeImageGenerationConfig(BaseImageGenerationConfig): if k == "size": # Convert "WxH" → "W*H" mapped["size"] = OPENAI_TO_DASHSCOPE_SIZE.get(v, v.replace("x", "*")) - elif k == "n": - mapped["image_count"] = v + else: + mapped[k] = v return mapped def get_complete_url( @@ -95,7 +99,10 @@ class DashScopeImageGenerationConfig(BaseImageGenerationConfig): litellm_params: dict, stream: bool | None = None, ) -> str: - return api_base or get_secret_str("DASHSCOPE_API_BASE_IMAGE") or DEFAULT_API_BASE + image_api_base: Final = ( + api_base if api_base and not api_base.rstrip("/").endswith(CHAT_COMPATIBLE_MODE_PATH) else None + ) + return image_api_base or get_secret_str("DASHSCOPE_API_BASE_IMAGE") or DEFAULT_API_BASE def validate_environment( self, diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index dd367e875de..2ecc179adac 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -14647,6 +14647,22 @@ "/v1/images/generations" ] }, + "dashscope/qwen-image-3.0": { + "litellm_provider": "dashscope", + "mode": "image_generation", + "source": "https://www.alibabacloud.com/help/en/model-studio/models", + "supported_endpoints": [ + "/v1/images/generations" + ] + }, + "dashscope/qwen-image-3.0-pro": { + "litellm_provider": "dashscope", + "mode": "image_generation", + "source": "https://www.alibabacloud.com/help/en/model-studio/models", + "supported_endpoints": [ + "/v1/images/generations" + ] + }, "databricks/databricks-bge-large-en": { "cache_creation_input_token_cost": 1.0003e-07, "cache_read_input_token_cost": 1.0003e-07, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index dd367e875de..2ecc179adac 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -14647,6 +14647,22 @@ "/v1/images/generations" ] }, + "dashscope/qwen-image-3.0": { + "litellm_provider": "dashscope", + "mode": "image_generation", + "source": "https://www.alibabacloud.com/help/en/model-studio/models", + "supported_endpoints": [ + "/v1/images/generations" + ] + }, + "dashscope/qwen-image-3.0-pro": { + "litellm_provider": "dashscope", + "mode": "image_generation", + "source": "https://www.alibabacloud.com/help/en/model-studio/models", + "supported_endpoints": [ + "/v1/images/generations" + ] + }, "databricks/databricks-bge-large-en": { "cache_creation_input_token_cost": 1.0003e-07, "cache_read_input_token_cost": 1.0003e-07, diff --git a/tests/test_litellm/test_dashscope_image_generation.py b/tests/test_litellm/test_dashscope_image_generation.py index c9f0df4febb..119efa010e0 100644 --- a/tests/test_litellm/test_dashscope_image_generation.py +++ b/tests/test_litellm/test_dashscope_image_generation.py @@ -1,5 +1,6 @@ """ -Unit tests for DashScope image generation support (qwen-image-2.0, qwen-image-2.0-pro). +Unit tests for DashScope image generation support (qwen-image-2.0, qwen-image-2.0-pro, +qwen-image-3.0, qwen-image-3.0-pro). Run in docker: pytest tests/test_litellm/test_dashscope_image_generation.py -v """ @@ -30,6 +31,8 @@ from litellm.llms.base_llm.chat.transformation import BaseLLMException [ "dashscope/qwen-image-2.0", "dashscope/qwen-image-2.0-pro", + "dashscope/qwen-image-3.0", + "dashscope/qwen-image-3.0-pro", ], ) def test_get_llm_provider_returns_dashscope(model_string: str): @@ -48,6 +51,8 @@ def test_get_llm_provider_returns_dashscope(model_string: str): [ ("dashscope/qwen-image-2.0", "dashscope"), ("dashscope/qwen-image-2.0-pro", "dashscope"), + ("dashscope/qwen-image-3.0", "dashscope"), + ("dashscope/qwen-image-3.0-pro", "dashscope"), ], ) def test_get_model_info_mode_is_image_generation( @@ -93,6 +98,19 @@ class TestDashScopeImageGenerationConfig: url = self.cfg.get_complete_url(custom, None, "qwen-image-2.0", {}, {}) assert url == custom + @pytest.mark.parametrize( + "chat_api_base", + [ + "https://dashscope.aliyuncs.com/compatible-mode/v1", + "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/", + ], + ) + def test_get_complete_url_ignores_chat_compatible_mode_base( + self, chat_api_base: str + ): + url = self.cfg.get_complete_url(chat_api_base, None, "qwen-image-3.0", {}, {}) + assert url == DEFAULT_API_BASE + def test_validate_environment_sets_auth_header(self): headers = self.cfg.validate_environment( headers={}, @@ -135,6 +153,27 @@ class TestDashScopeImageGenerationConfig: assert messages[0]["content"][0]["text"] == "a puppy on green grass" assert req["parameters"]["size"] == "1024*1024" + @pytest.mark.parametrize("model", ["qwen-image-3.0", "qwen-image-3.0-pro"]) + def test_transform_request_qwen_image_3(self, model: str): + req = self.cfg.transform_image_generation_request( + model=model, + prompt="a poster with small multilingual text", + optional_params=self.cfg.map_openai_params( + non_default_params={"size": "2048x2048", "n": 6}, + optional_params={}, + model=model, + drop_params=False, + ), + litellm_params={}, + headers={}, + ) + assert req["model"] == model + assert req["input"]["messages"][0]["content"][0]["text"] == ( + "a poster with small multilingual text" + ) + assert req["parameters"]["size"] == "2048*2048" + assert req["parameters"]["n"] == 6 + def test_transform_request_empty_params(self): req = self.cfg.transform_image_generation_request( model="qwen-image-2.0-pro", @@ -238,6 +277,48 @@ class TestDashScopeImageGenerationConfig: assert result.data[0].url == "https://example.com/img1.png" assert result.data[1].url == "https://example.com/img2.png" + def test_transform_response_multiple_images_in_one_choice(self): + body = { + "output": { + "choices": [ + { + "finish_reason": "stop", + "message": { + "role": "assistant", + "content": [ + {"image": "https://example.com/img1.png", "type": "image"}, + {"image": "https://example.com/img2.png", "type": "image"}, + ], + }, + } + ] + }, + "usage": { + "output_width": 1024, + "output_height": 1024, + "output_image_count": 2, + }, + } + mock_resp = MagicMock(spec=httpx.Response) + mock_resp.status_code = 200 + mock_resp.headers = {} + mock_resp.json.return_value = body + + result = self.cfg.transform_image_generation_response( + model="qwen-image-3.0", + raw_response=mock_resp, + model_response=ImageResponse(), + logging_obj=MagicMock(), + request_data={}, + optional_params={}, + litellm_params={}, + encoding=None, + ) + assert [image.url for image in result.data] == [ + "https://example.com/img1.png", + "https://example.com/img2.png", + ] + def test_transform_response_raises_on_non_200_status(self): mock_resp = MagicMock(spec=httpx.Response) mock_resp.status_code = 400 @@ -294,14 +375,14 @@ class TestDashScopeImageGenerationConfig: ) assert mapped["size"] == "1024*1024" - def test_map_openai_params_n_to_image_count(self): + def test_map_openai_params_n_passthrough(self): mapped = self.cfg.map_openai_params( non_default_params={"n": 2}, optional_params={}, model="qwen-image-2.0", drop_params=False, ) - assert mapped["image_count"] == 2 + assert mapped == {"n": 2} def test_map_openai_params_unknown_size_uses_asterisk(self): mapped = self.cfg.map_openai_params( @@ -338,7 +419,15 @@ class TestDashScopeImageGenerationConfig: # --------------------------------------------------------------------------- -def test_litellm_image_generation_dashscope_end_to_end(): +@pytest.mark.parametrize( + "model", + [ + "dashscope/qwen-image-2.0", + "dashscope/qwen-image-3.0", + "dashscope/qwen-image-3.0-pro", + ], +) +def test_litellm_image_generation_dashscope_end_to_end(model: str): mock_response_body = { "output": { "choices": [ @@ -374,7 +463,7 @@ def test_litellm_image_generation_dashscope_end_to_end(): mock_post.return_value = mock_http_response response = litellm.image_generation( - model="dashscope/qwen-image-2.0", + model=model, prompt="a puppy playing on green grass", api_key="sk-test-key", size="1024x1024", @@ -392,7 +481,7 @@ def test_litellm_image_generation_dashscope_end_to_end(): called_url = ( call_args[0][0] if call_args[0] else call_args.kwargs.get("url", "") ) - assert "dashscope" in called_url or "aliyuncs" in called_url + assert called_url == DEFAULT_API_BASE # Verify request body contains DashScope format call_kwargs = call_args[1] if call_args[1] else {} @@ -400,3 +489,4 @@ def test_litellm_image_generation_dashscope_end_to_end(): body = call_kwargs["json"] assert "input" in body assert "messages" in body["input"] + assert body["parameters"]["size"] == "1024*1024" From 6d7fafa3470541d797077d8e7854000d7c7f4c79 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 20:37:11 -0700 Subject: [PATCH 23/44] fix(anthropic): close hybrid tool-name allowlist gap and keep native tools through guardrails --- .../chat/guardrail_translation/handler.py | 21 +++++++++++---- .../adapters/transformation.py | 4 +-- .../base_llm/guardrail_translation/utils.py | 19 +++++++------ .../test_anthropic_guardrail_handler.py | 18 +++++++++++++ .../proxy/test_tools_allowlist_enforcement.py | 27 +++++++++++++++++++ 5 files changed, 74 insertions(+), 15 deletions(-) diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py index 2afea1444f9..b9ca18c7843 100644 --- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py +++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py @@ -24,10 +24,12 @@ from litellm._logging import verbose_proxy_logger from litellm.llms.anthropic.chat.transformation import AnthropicConfig from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import ( LiteLLMAnthropicMessagesAdapter, + is_provider_native_tool_dict, ) from litellm.llms.base_llm.guardrail_translation.base_translation import BaseTranslation from litellm.llms.base_llm.guardrail_translation.utils import ( anthropic_tool_name, + anthropic_tool_names, effective_scan_only_tool_results_for_guardrail, effective_skip_system_message_for_guardrail, effective_skip_tool_message_for_guardrail, @@ -360,7 +362,13 @@ class AnthropicMessagesHandler(BaseTranslation): structured_messages: Final = [full_structured_messages[index] for index in scoped_message_indices] tools_to_check: Final[list[ChatCompletionToolParam]] = ( - [] if scan_only_tool_results else chat_completion_compatible_request.get("tools", []) + [] + if scan_only_tool_results + else [ + tool + for tool in chat_completion_compatible_request.get("tools", []) + if not is_provider_native_tool_dict(tool) + ] ) # Step 1: Extract all text content and images @@ -419,7 +427,10 @@ class AnthropicMessagesHandler(BaseTranslation): tool_name=anthropic_tool_name, ) if scan_only_tool_results - else anthropic_tools + else [ + *(tool for tool in data.get("tools") or [] if is_provider_native_tool_dict(tool)), + *anthropic_tools, + ] ) guardrailed_structured_messages: Final = guardrailed_inputs.get("structured_messages") @@ -677,9 +688,9 @@ class AnthropicMessagesHandler(BaseTranslation): ) def extract_request_tool_names(self, data: dict) -> list[str]: - """Extract tool names from Anthropic messages request (tools[].name, or - tools[].function.name for OpenAI-format tools the bridge forwards verbatim).""" - return [name for tool in data.get("tools") or [] if (name := anthropic_tool_name(tool))] + """Extract every tool name in an Anthropic messages request: tools[].name, plus + tools[].function.name for OpenAI-format tools the bridge forwards verbatim.""" + return [name for tool in data.get("tools") or [] for name in anthropic_tool_names(tool)] @classmethod def _extract_input_text_and_images( diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index ca90df2ff75..c958f9f242f 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -27,7 +27,7 @@ def _is_openai_function_tool(tool: Mapping[str, object]) -> bool: return tool.get("type") == "function" and "function" in tool -def _is_provider_native_tool_dict(tool: Mapping[str, object]) -> bool: +def is_provider_native_tool_dict(tool: Mapping[str, object]) -> bool: if len(tool) != 1: return False key, value = next(iter(tool.items())) @@ -786,7 +786,7 @@ class LiteLLMAnthropicMessagesAdapter: new_tools.append(tool) continue - if _is_openai_function_tool(tool) or _is_provider_native_tool_dict(tool): + if _is_openai_function_tool(tool) or is_provider_native_tool_dict(tool): new_tools.append(cast(ChatCompletionToolParam, tool)) # cast-ok: passed through verbatim to provider continue diff --git a/litellm/llms/base_llm/guardrail_translation/utils.py b/litellm/llms/base_llm/guardrail_translation/utils.py index 4494f8451f4..aefe3861e3c 100644 --- a/litellm/llms/base_llm/guardrail_translation/utils.py +++ b/litellm/llms/base_llm/guardrail_translation/utils.py @@ -209,17 +209,20 @@ def openai_tool_name(tool: object) -> str | None: return flat_name if isinstance(flat_name, str) else None -def anthropic_tool_name(tool: object) -> str | None: - """Anthropic tools carry a flat ``name``; an OpenAI-format function tool, which the - non-Anthropic bridge forwards verbatim, carries it under ``function.name`` instead.""" +def anthropic_tool_names(tool: object) -> tuple[str, ...]: + """Every name a /v1/messages tool dict can act under: the flat Anthropic ``name`` plus + ``function.name`` for OpenAI-format tools the bridge forwards verbatim. Allowlist checks + must see both, or a decoy flat name could smuggle a disallowed ``function.name`` through.""" if not isinstance(tool, dict): - return None - flat_name: Final = tool.get("name") - if isinstance(flat_name, str): - return flat_name + return () function: Final = tool.get("function") if tool.get("type") == "function" else None function_name: Final = function.get("name") if isinstance(function, dict) else None - return function_name if isinstance(function_name, str) else None + return tuple(name for name in (tool.get("name"), function_name) if isinstance(name, str) and name) + + +def anthropic_tool_name(tool: object) -> str | None: + names: Final = anthropic_tool_names(tool) + return names[0] if names else None def merge_returned_tools_into_request_tools( diff --git a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py index 2b392456763..46bdd855a28 100644 --- a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py +++ b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py @@ -290,6 +290,24 @@ class TestAnthropicMessagesHandlerInputProcessing: assert data.get("litellm_metadata", {}).get("guardrails") assert guardrail.dynamic_params == {"policy_id": "policy-123"} + @pytest.mark.asyncio + async def test_provider_native_tools_survive_guardrail_round_trip(self): + handler = AnthropicMessagesHandler() + guardrail = MockPassThroughGuardrail(guardrail_name="test") + data = { + "model": "gemini-2.5-flash", + "messages": [{"role": "user", "content": "coffee shops near Union Square?"}], + "tools": [ + {"googleMaps": {"enable_widget": True}}, + {"name": "get_weather", "input_schema": {"type": "object", "properties": {}}}, + ], + } + + await handler.process_input_messages(data=data, guardrail_to_apply=guardrail) + + assert {"googleMaps": {"enable_widget": True}} in data["tools"] + assert [tool["name"] for tool in data["tools"] if "name" in tool] == ["get_weather"] + @pytest.mark.asyncio async def test_midturn_system_correction_is_guardrailed_when_top_level_system_is_skipped( self, diff --git a/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py b/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py index c881effada5..3e9c7c14b95 100644 --- a/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py +++ b/tests/test_litellm/proxy/test_tools_allowlist_enforcement.py @@ -106,6 +106,19 @@ class TestExtractRequestToolNames: "run_sql", ] + def test_anthropic_hybrid_tool_yields_every_name(self): + data = { + "tools": [ + {"type": "function", "name": "decoy", "function": {"name": "blocked_fn"}}, + {"type": "function", "name": "", "function": {"name": "hidden_fn"}}, + ] + } + assert extract_request_tool_names("/v1/messages", data) == [ + "decoy", + "blocked_fn", + "hidden_fn", + ] + def test_generate_content_tools(self): data = { "tools": [ @@ -186,6 +199,20 @@ class TestCheckToolsAllowlist: assert exc_info.value.type == ProxyErrorTypes.tool_access_denied assert "get_weather" in str(exc_info.value.message) + @pytest.mark.asyncio + async def test_hybrid_tool_with_decoy_name_raises_on_messages_route(self): + token = _token(metadata={"allowed_tools": ["decoy"]}) + body = {"tools": [{"type": "function", "name": "decoy", "function": {"name": "run_sql"}}]} + with pytest.raises(ProxyException) as exc_info: + await check_tools_allowlist( + request_body=body, + valid_token=token, + team_object=None, + route="/v1/messages", + ) + assert exc_info.value.type == ProxyErrorTypes.tool_access_denied + assert "run_sql" in str(exc_info.value.message) + @pytest.mark.asyncio async def test_disallowed_custom_tool_raises_on_responses_route(self): token = _token(metadata={"allowed_tools": ["other_tool"]}) From 449c0913919f6e7f06d84b2504af8c2036328933 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 21:10:12 -0700 Subject: [PATCH 24/44] fix(realtime): carry audio output tokens into response.done usage so Gemini Live native audio bills at the audio rate --- .../transformation.py | 1 + litellm/types/llms/openai.py | 2 + .../test_gemini_realtime_transformation.py | 50 +++++++++++++++++++ 3 files changed, 53 insertions(+) diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py index 1381ad36d12..f39df38d069 100644 --- a/litellm/responses/litellm_completion_transformation/transformation.py +++ b/litellm/responses/litellm_completion_transformation/transformation.py @@ -2661,6 +2661,7 @@ class LiteLLMCompletionResponsesConfig: optional_output_details: Final[dict[str, int]] = { field: value for field, value in ( + ("audio_tokens", getattr(completion_details, "audio_tokens", None)), ("text_tokens", getattr(completion_details, "text_tokens", None)), ("image_tokens", getattr(completion_details, "image_tokens", None)), ) diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index 45f6b5c55a9..62a8e34c668 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -1258,6 +1258,8 @@ class ResponsesAPIRequestParams(ResponsesAPIOptionalRequestParams, total=False): class OutputTokensDetails(BaseLiteLLMOpenAIResponseObject): + audio_tokens: int | None = None + reasoning_tokens: int | None = None text_tokens: int | None = None diff --git a/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py b/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py index 42e330925a0..a67c8d32c5d 100644 --- a/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py +++ b/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py @@ -1864,3 +1864,53 @@ def test_map_openai_params_drops_stock_voice_case_insensitively(): passthrough = cfg.map_openai_params(optional_params={}, non_default_params={"voice": "Kore"}) assert passthrough["generationConfig"]["speechConfig"]["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == "Kore" + + +def test_gemini_response_done_bills_audio_output_tokens_at_audio_rate(monkeypatch): + """Regression for the Gemini Live AUDIO output breakdown: responseTokensDetails + must survive into response.done usage and bill at output_cost_per_audio_token, + not the text rate.""" + from litellm.cost_calculator import ( + RealtimeAPITokenUsageProcessor, + handle_realtime_stream_cost_calculation, + ) + + monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url="")) + + config = GeminiRealtimeConfig() + done_event = config.transform_response_done_event( + message={ + "serverContent": {"turnComplete": True}, + "usageMetadata": { + "promptTokenCount": 377, + "responseTokenCount": 51, + "totalTokenCount": 428, + "promptTokensDetails": [{"modality": "TEXT", "tokenCount": 377}], + "responseTokensDetails": [{"modality": "AUDIO", "tokenCount": 51}], + "thoughtsTokenCount": 37, + }, + }, + current_response_id="resp_lit6277", + current_conversation_id="conv_lit6277", + output_items=None, + ) + + usage = done_event["response"]["usage"] + assert usage["output_tokens_details"]["audio_tokens"] == 51 + assert usage["output_token_details"]["audio_tokens"] == 51 + + results = [done_event] + combined_usage = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results( + results=results, + ) + assert combined_usage.completion_tokens_details is not None + assert combined_usage.completion_tokens_details.audio_tokens == 51 + + cost = handle_realtime_stream_cost_calculation( + results=results, + combined_usage_object=combined_usage, + custom_llm_provider="gemini", + litellm_model_name="gemini-2.5-flash-native-audio-preview-12-2025", + ) + assert cost == pytest.approx(377 * 5e-07 + 51 * 1.2e-05 + 37 * 2e-06) From f7228a46702b19e384626b25a0fa84d6d7cb5ed9 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 21:47:15 -0700 Subject: [PATCH 25/44] fix(streaming): preserve parsed-chunk provider_specific_fields so Vertex flex streams bill at flex rates --- .../litellm_core_utils/streaming_handler.py | 24 +++++++- .../test_streaming_handler.py | 59 +++++++++++++++++++ 2 files changed, 80 insertions(+), 3 deletions(-) diff --git a/litellm/litellm_core_utils/streaming_handler.py b/litellm/litellm_core_utils/streaming_handler.py index f6340426c1b..0f46f1b718c 100644 --- a/litellm/litellm_core_utils/streaming_handler.py +++ b/litellm/litellm_core_utils/streaming_handler.py @@ -8,11 +8,12 @@ import time import traceback from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence from dataclasses import dataclass +from types import MappingProxyType from typing import Any, Final, NoReturn, Protocol, TypeVar, cast import anyio import httpx -from pydantic import BaseModel +from pydantic import BaseModel, ValidationError from typing_extensions import NotRequired, TypedDict import litellm @@ -182,6 +183,23 @@ class _VertexChunkLike(Protocol): candidates: Sequence[_VertexCandidateLike] +class _ParsedChunkHiddenParams(BaseModel): + provider_specific_fields: Mapping[str, object] | None = None + + +def _provider_hidden_params(chunk: object) -> Mapping[str, object] | None: + hidden: Final[object] = getattr(chunk, "_hidden_params", None) + if not isinstance(hidden, dict): + return None + try: + parsed: Final = _ParsedChunkHiddenParams.model_validate(hidden) + except ValidationError: + return None + if not parsed.provider_specific_fields: + return None + return MappingProxyType({"provider_specific_fields": dict(parsed.provider_specific_fields)}) + + class CustomStreamWrapper: def __init__( self, @@ -801,7 +819,7 @@ class CustomStreamWrapper: except Exception as e: raise e - def model_response_creator(self, chunk: dict | None = None, hidden_params: dict | None = None): + def model_response_creator(self, chunk: dict | None = None, hidden_params: Mapping[str, object] | None = None): _model: Final = self._cached_model_name _logging_obj_llm_provider: Final = self._cached_logging_llm_provider @@ -1504,7 +1522,7 @@ class CustomStreamWrapper: def chunk_creator(self, chunk: Any): if hasattr(chunk, "id"): self.response_id = chunk.id - model_response = self.model_response_creator() + model_response = self.model_response_creator(hidden_params=_provider_hidden_params(chunk)) response_obj: dict[str, Any] = {} try: # return this for all models diff --git a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py index b5e33a4e421..d39c84ce5e7 100644 --- a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py +++ b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py @@ -4460,3 +4460,62 @@ def test_handle_stream_fallback_error_restores_context_only_after_exception_mapp finally: trace_id_var.set("") session_id_var.set("") + + +def test_chunk_creator_preserves_hidden_provider_specific_fields_from_parsed_chunk(): + """ + Vertex/Gemini chunk_parser stores usageMetadata.trafficType in the parsed + chunk's _hidden_params["provider_specific_fields"], but chunk_creator builds + a fresh ModelResponseStream per outgoing chunk. Before the fix those hidden + provider fields were dropped, so streaming flex traffic was billed at + standard rates (LIT-6292). + """ + wrapper = CustomStreamWrapper( + completion_stream=None, + model="gemini-3.5-flash", + logging_obj=MagicMock(), + custom_llm_provider="vertex_ai", + ) + parsed_chunk = ModelResponseStream( + choices=[StreamingChoices(index=0, delta=Delta(content="hello", role="assistant"), finish_reason=None)], + ) + parsed_chunk._hidden_params["provider_specific_fields"] = {"traffic_type": "ON_DEMAND_FLEX"} + + result = wrapper.chunk_creator(chunk=parsed_chunk) + + assert result is not None + assert result._hidden_params["provider_specific_fields"] == {"traffic_type": "ON_DEMAND_FLEX"} + + +@pytest.mark.asyncio +async def test_async_stream_assembled_response_keeps_vertex_traffic_type(logging_obj: Logging): + """ + End-to-end through the async wrapper: the assembled response handed to cost + tracking must carry traffic_type so flex/priority tiers price correctly. + """ + content_chunk = ModelResponseStream( + choices=[StreamingChoices(index=0, delta=Delta(content="hello", role="assistant"), finish_reason=None)], + ) + final_chunk = ModelResponseStream( + choices=[StreamingChoices(index=0, delta=Delta(content=""), finish_reason="stop")], + ) + setattr(final_chunk, "usage", Usage(prompt_tokens=7, completion_tokens=5, total_tokens=12)) + final_chunk._hidden_params["provider_specific_fields"] = {"traffic_type": "ON_DEMAND_FLEX"} + + async def _stream(): + yield content_chunk + yield final_chunk + + wrapper = CustomStreamWrapper( + completion_stream=_stream(), + model="gemini-3.5-flash", + logging_obj=logging_obj, + custom_llm_provider="vertex_ai", + stream_options={"include_usage": True}, + ) + + received = [chunk async for chunk in wrapper] + + assembled = litellm.stream_chunk_builder(chunks=received, messages=[{"role": "user", "content": "hi"}]) + assert assembled is not None + assert assembled._hidden_params["provider_specific_fields"]["traffic_type"] == "ON_DEMAND_FLEX" From ab71807985237352965250bc3c3e9be73ef803da Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 22:23:17 -0700 Subject: [PATCH 26/44] test(streaming): drop redundant docstrings from flex-tier regression tests --- .../litellm_core_utils/test_streaming_handler.py | 11 ----------- 1 file changed, 11 deletions(-) diff --git a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py index d39c84ce5e7..5329edce47e 100644 --- a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py +++ b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py @@ -4463,13 +4463,6 @@ def test_handle_stream_fallback_error_restores_context_only_after_exception_mapp def test_chunk_creator_preserves_hidden_provider_specific_fields_from_parsed_chunk(): - """ - Vertex/Gemini chunk_parser stores usageMetadata.trafficType in the parsed - chunk's _hidden_params["provider_specific_fields"], but chunk_creator builds - a fresh ModelResponseStream per outgoing chunk. Before the fix those hidden - provider fields were dropped, so streaming flex traffic was billed at - standard rates (LIT-6292). - """ wrapper = CustomStreamWrapper( completion_stream=None, model="gemini-3.5-flash", @@ -4489,10 +4482,6 @@ def test_chunk_creator_preserves_hidden_provider_specific_fields_from_parsed_chu @pytest.mark.asyncio async def test_async_stream_assembled_response_keeps_vertex_traffic_type(logging_obj: Logging): - """ - End-to-end through the async wrapper: the assembled response handed to cost - tracking must carry traffic_type so flex/priority tiers price correctly. - """ content_chunk = ModelResponseStream( choices=[StreamingChoices(index=0, delta=Delta(content="hello", role="assistant"), finish_reason=None)], ) From 99af9ad9ebe0169eddee5347617ae9c088524421 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 22:35:46 -0700 Subject: [PATCH 27/44] fix(anthropic): carry tool_reference tool results through the guardrail translation round trip --- basedpyright-code-budget.json | 14 +- .../transformation.py | 35 +++-- .../anthropic_cache_control_hook.py | 2 +- .../prompt_templates/factory.py | 17 ++- .../adapters/transformation.py | 132 ++++++------------ litellm/llms/gemini/chat/transformation.py | 21 ++- litellm/llms/mistral/chat/transformation.py | 2 +- litellm/types/llms/anthropic.py | 7 +- litellm/types/llms/openai.py | 16 ++- ...responses_transformation_transformation.py | 36 +++++ ...llm_core_utils_prompt_templates_factory.py | 49 +++++++ .../test_anthropic_guardrail_handler.py | 69 +++++++++ ...al_pass_through_adapters_transformation.py | 69 +++++++++ type-discipline-budget.json | 4 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 15 +- 15 files changed, 363 insertions(+), 125 deletions(-) diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json index 3357212a6c8..e768e83c899 100644 --- a/basedpyright-code-budget.json +++ b/basedpyright-code-budget.json @@ -6,10 +6,10 @@ "limit": 2564 }, "reportAssignmentType": { - "limit": 320 + "limit": 319 }, "reportAttributeAccessIssue": { - "limit": 483 + "limit": 480 }, "reportCallIssue": { "limit": 113 @@ -30,7 +30,7 @@ "limit": 7 }, "reportGeneralTypeIssues": { - "limit": 154 + "limit": 105 }, "reportIncompatibleMethodOverride": { "limit": 56 @@ -99,19 +99,19 @@ "limit": 0 }, "reportUnknownArgumentType": { - "limit": 44528 + "limit": 44526 }, "reportUnknownLambdaType": { "limit": 109 }, "reportUnknownMemberType": { - "limit": 38804 + "limit": 38782 }, "reportUnknownParameterType": { "limit": 19829 }, "reportUnknownVariableType": { - "limit": 30355 + "limit": 30349 }, "reportUnnecessaryCast": { "limit": 117 @@ -123,7 +123,7 @@ "limit": 5 }, "reportUnnecessaryIsInstance": { - "limit": 833 + "limit": 831 }, "reportUntypedBaseClass": { "limit": 0 diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py index 17815976b4a..85fb0bc8dc6 100644 --- a/litellm/completion_extras/litellm_responses_transformation/transformation.py +++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py @@ -59,9 +59,11 @@ if TYPE_CHECKING: from litellm.types.llms.openai import ( ALL_RESPONSES_API_TOOL_PARAMS, AllMessageValues, + ChatCompletionFileObject, ChatCompletionImageObject, ChatCompletionRedactedThinkingBlock, ChatCompletionThinkingBlock, + ChatCompletionToolReferenceObject, OpenAIMessageContentListBlock, ) from litellm.types.utils import Choices @@ -175,6 +177,16 @@ def _map_incomplete_reason_to_finish_reason(incomplete_reason: str | None) -> Li return "length" +def _input_file_from_file_value(file_value: object) -> dict[str, object]: + if not isinstance(file_value, dict): + return {"type": "input_file"} + file_dict: Final = cast("dict[str, object]", file_value) # cast-ok: runtime dict checked + return { + "type": "input_file", + **{key: file_dict[key] for key in ("file_id", "file_data", "filename") if key in file_dict}, + } + + def _incomplete_reason_from_response_payload(response_payload: object) -> str | None: if not isinstance(response_payload, Mapping): return None @@ -957,7 +969,12 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): content: str | list[object] | Iterable[ - Union["OpenAIMessageContentListBlock", "ChatCompletionThinkingBlock", "ChatCompletionRedactedThinkingBlock"] + Union[ + "OpenAIMessageContentListBlock", + "ChatCompletionThinkingBlock", + "ChatCompletionRedactedThinkingBlock", + "ChatCompletionToolReferenceObject", + ] ] | None, role: str, @@ -1006,17 +1023,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): result.append(converted) verbose_logger.debug("Chat provider: image -> %s", converted) elif item_type == "file": - # Map Chat Completion file to Responses API input_file - # {"type": "file", "file": {"file_data": "...", "filename": "..."}} - # -> {"type": "input_file", "file_data": "...", "filename": "..."} - file_data = item.get("file", {}) - converted = {"type": "input_file"} - if isinstance(file_data, dict): - for key in ["file_id", "file_data", "filename"]: - if key in file_data: - converted[key] = file_data[key] + converted = _input_file_from_file_value( + cast("ChatCompletionFileObject", item).get("file"), # cast-ok: type tag checked + ) result.append(converted) verbose_logger.debug("Chat provider: file -> %s", converted) + elif item_type == "tool_reference": + verbose_logger.debug( + "Chat provider: tool_reference has no responses API equivalent; skipped" + ) elif item_type in [ "input_text", "input_image", diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py index ef2edbf1007..f972bad47e6 100644 --- a/litellm/integrations/anthropic_cache_control_hook.py +++ b/litellm/integrations/anthropic_cache_control_hook.py @@ -376,7 +376,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): # 2. list of objects - only apply to last item per Anthropic spec elif isinstance(message_content, list): if len(message_content) > 0 and isinstance(message_content[-1], dict): - message_content[-1]["cache_control"] = control + message_content[-1]["cache_control"] = control # pyright: ignore[reportGeneralTypeIssues] # loose runtime dict return message @staticmethod diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index 86cfbf70255..795fb36961e 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -1412,7 +1412,7 @@ def convert_to_gemini_tool_call_result( ) except Exception as e: verbose_logger.warning("Failed to process image in tool response: %s", e) - elif content_type in ("file", "input_file"): + elif content_type in ("file", "input_file"): # pyright: ignore[reportUnnecessaryContains] # loose runtime dict # Extract file for inline_data (for tool results with PDF, audio, video, etc.) file_data = content.get("file_data", "") if not file_data: @@ -1564,14 +1564,23 @@ def convert_to_anthropic_tool_result( } """ anthropic_content: ( - str | list[AnthropicMessagesToolResultContent | AnthropicMessagesImageParam | AnthropicMessagesDocumentParam] + str + | list[ + AnthropicMessagesToolResultContent + | AnthropicMessagesImageParam + | AnthropicMessagesDocumentParam + | ToolReference + ] ) = "" if isinstance(message["content"], str): anthropic_content = message["content"] elif isinstance(message["content"], list): content_list: Final = message["content"] anthropic_content_list: list[ - AnthropicMessagesToolResultContent | AnthropicMessagesImageParam | AnthropicMessagesDocumentParam + AnthropicMessagesToolResultContent + | AnthropicMessagesImageParam + | AnthropicMessagesDocumentParam + | ToolReference ] = [] for content in content_list: if content["type"] == "text": @@ -1614,6 +1623,8 @@ def convert_to_anthropic_tool_result( original_content_element=content, ) anthropic_content_list.append(cast(AnthropicMessagesImageParam, _anthropic_image_param)) + elif content["type"] == "tool_reference": + anthropic_content_list.append(ToolReference(type="tool_reference", tool_name=content["tool_name"])) elif content["type"] == "file": file_content = cast(ChatCompletionFileObject, content) _file_block = anthropic_process_openai_file_message(file_content) diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index 109017bda27..0167dc73493 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -1,8 +1,8 @@ import copy import hashlib import json -from collections.abc import AsyncIterator, Iterator, Mapping -from typing import TYPE_CHECKING, Any, Final, Literal, TypeVar, cast +from collections.abc import AsyncIterator, Iterator, Mapping, Sequence +from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias, TypeVar, cast import litellm from litellm.llms.anthropic.experimental_pass_through.utils import ( @@ -125,7 +125,9 @@ from litellm.types.llms.openai import ( ChatCompletionToolMessage, ChatCompletionToolParam, ChatCompletionToolParamFunctionChunk, + ChatCompletionToolReferenceObject, ChatCompletionUserMessage, + ToolMessageContentPart, ) from litellm.types.utils import Choices, ModelResponse, StreamingChoices, Usage @@ -134,6 +136,8 @@ from .streaming_iterator import AnthropicStreamWrapper if TYPE_CHECKING: from litellm.types.llms.anthropic import ContentBlockContentBlockDict +ToolResultContent: TypeAlias = str | list[ToolMessageContentPart] + class AnthropicAdapter: def __init__(self) -> None: @@ -411,90 +415,13 @@ class LiteLLMAnthropicMessagesAdapter: self._add_cache_control_if_applicable(content, doc_obj, model) new_user_content_list.append(doc_obj) elif content.get("type") == "tool_result": - if "content" not in content: - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content="", - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) - elif isinstance(content.get("content"), str): - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content=str(content.get("content", "")), - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) - elif isinstance(content.get("content"), list): - # Combine all content items into a single tool message - # to avoid creating multiple tool_result blocks with the same ID - # (each tool_use must have exactly one tool_result) - content_items = list(content.get("content", [])) - - # Single-item text keeps the backward-compatible string format; a single - # image or document becomes a structured image_url part - if len(content_items) == 1: - c = content_items[0] - if isinstance(c, str): - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content=c, - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) - elif isinstance(c, dict): - if c.get("type") == "text": - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content=c.get("text", ""), - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) - elif c.get("type") in ("image", "document"): - image_part = self._tool_result_image_part(c.get("source")) - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content=[image_part] # mutable-ok: content must be a json list - if image_part - else "", - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) - else: - # For multiple content items, combine into a single tool message - # with list content to preserve all items while having one tool_use_id - combined_content_parts: list[ - ChatCompletionTextObject | ChatCompletionImageObject - ] = [] - for c in content_items: - if isinstance(c, str): - combined_content_parts.append(ChatCompletionTextObject(type="text", text=c)) - elif isinstance(c, dict): - if c.get("type") == "text": - combined_content_parts.append( - ChatCompletionTextObject( - type="text", - text=c.get("text", ""), - ) - ) - elif c.get("type") in ("image", "document"): - image_part = self._tool_result_image_part(c.get("source")) - if image_part: - combined_content_parts.append(image_part) - # Create a single tool message with combined content - if combined_content_parts: - tool_result = ChatCompletionToolMessage( - role="tool", - tool_call_id=content.get("tool_use_id", ""), - content=combined_content_parts, - ) - self._add_cache_control_if_applicable(content, tool_result, model) - tool_message_list.append(tool_result) + tool_result = ChatCompletionToolMessage( + role="tool", + tool_call_id=content.get("tool_use_id", ""), + content=self._tool_result_content(content.get("content")), + ) + self._add_cache_control_if_applicable(content, tool_result, model) + tool_message_list.append(tool_result) if len(tool_message_list) > 0: new_messages.extend(tool_message_list) @@ -1209,6 +1136,39 @@ class LiteLLMAnthropicMessagesAdapter: return None + def _tool_result_content(self, raw_content: object) -> ToolResultContent: + if isinstance(raw_content, str): + return raw_content + if not isinstance(raw_content, list): + return "" + items: Final = cast(Sequence[object], raw_content) # cast-ok: untrusted client payload + parts: Final = tuple(part for part in (self._tool_result_part(item) for item in items) if part is not None) + match parts: + case (): + return "" + case ({"type": "text", "text": str(text)},): + return text + case _: + return list(parts) # mutable-ok: content must be a json list + + def _tool_result_part(self, item: object) -> ToolMessageContentPart | None: + if isinstance(item, str): + return ChatCompletionTextObject(type="text", text=item) + if not isinstance(item, dict): + return None + block: Final = cast(Mapping[str, object], item) # cast-ok: untrusted client payload + match block.get("type"): + case "text": + return ChatCompletionTextObject(type="text", text=str(block.get("text") or "")) + case "image" | "document": + return self._tool_result_image_part(block.get("source")) + case "tool_reference": + return ChatCompletionToolReferenceObject( + type="tool_reference", tool_name=str(block.get("tool_name") or "") + ) + case _: + return None + def _tool_result_image_part(self, image_source: object) -> ChatCompletionImageObject | None: if not isinstance(image_source, dict): return None diff --git a/litellm/llms/gemini/chat/transformation.py b/litellm/llms/gemini/chat/transformation.py index bc12995057e..1a67b33665b 100644 --- a/litellm/llms/gemini/chat/transformation.py +++ b/litellm/llms/gemini/chat/transformation.py @@ -8,7 +8,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import ( from litellm.litellm_core_utils.prompt_templates.image_handling import ( convert_url_to_base64, ) -from litellm.types.llms.openai import AllMessageValues, ChatCompletionFileObject +from litellm.types.llms.openai import AllMessageValues, ChatCompletionFileObject, ChatCompletionImageObject from litellm.types.llms.vertex_ai import ContentType, PartType from litellm.utils import supports_reasoning @@ -16,6 +16,13 @@ from ...vertex_ai.gemini.transformation import _gemini_convert_messages_with_his from ...vertex_ai.gemini.vertex_and_google_ai_studio_gemini import VertexGeminiConfig +def _image_url_fields(img_element: ChatCompletionImageObject) -> tuple[str | None, str | None, str | None]: + image_value: Final = img_element.get("image_url") + if isinstance(image_value, dict): + return image_value.get("url"), image_value.get("format"), image_value.get("detail") + return image_value, None, None + + class GoogleAIStudioGeminiConfig(VertexGeminiConfig): """ Reference: https://ai.google.dev/api/rest/v1beta/GenerationConfig @@ -118,16 +125,8 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig): _parts: list[PartType] = [] for element in _message_content: if element.get("type") == "image_url": - img_element = element - _image_url: str | None = None - format: str | None = None - detail: str | None = None - if isinstance(img_element.get("image_url"), dict): - _image_url = img_element["image_url"].get("url") - format = img_element["image_url"].get("format") - detail = img_element["image_url"].get("detail") - else: - _image_url = img_element.get("image_url") + img_element = cast(ChatCompletionImageObject, element) # cast-ok: runtime type tag checked + _image_url, format, detail = _image_url_fields(img_element) if _image_url and "https://" in _image_url: image_obj = convert_to_anthropic_image_obj(_image_url, format=format) converted_image_url = convert_generic_image_chunk_to_openai_image_obj(image_obj) diff --git a/litellm/llms/mistral/chat/transformation.py b/litellm/llms/mistral/chat/transformation.py index 0d9577669a4..0c95fd4df07 100644 --- a/litellm/llms/mistral/chat/transformation.py +++ b/litellm/llms/mistral/chat/transformation.py @@ -292,7 +292,7 @@ class MistralConfig(OpenAIGPTConfig): file_id = file_content.get("file", {}).get("file_id") if file_id: # Replace 'file' with 'file_id' - file_content["file_id"] = file_id + file_content["file_id"] = file_id # pyright: ignore[reportGeneralTypeIssues] # legacy in-place rewrite of the block shape file_content.pop("file", None) return messages diff --git a/litellm/types/llms/anthropic.py b/litellm/types/llms/anthropic.py index f127366cc21..eb6cf41285f 100644 --- a/litellm/types/llms/anthropic.py +++ b/litellm/types/llms/anthropic.py @@ -324,7 +324,12 @@ class AnthropicMessagesToolResultParam(TypedDict, total=False): is_error: bool content: ( str - | Iterable[AnthropicMessagesToolResultContent | AnthropicMessagesImageParam | AnthropicMessagesDocumentParam] + | Iterable[ + AnthropicMessagesToolResultContent + | AnthropicMessagesImageParam + | AnthropicMessagesDocumentParam + | ToolReference + ] ) cache_control: dict | ChatCompletionCachedContent | None diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index 45f6b5c55a9..beb788b64c3 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -1,7 +1,7 @@ from collections.abc import Iterable, Mapping from enum import Enum from os import PathLike -from typing import IO, Any, Final, Literal, Optional, Union +from typing import IO, Any, Final, Literal, Optional, TypeAlias, Union import httpx from openai import Omit @@ -820,9 +820,21 @@ class ChatCompletionAssistantMessage(OpenAIChatCompletionAssistantMessage, total reasoning_items: list[ChatCompletionReasoningItem] | None +class ChatCompletionToolReferenceObject(TypedDict): + """Anthropic tool-search result block, carried through untouched so it survives a round trip.""" + + type: Literal["tool_reference"] # writable-ok: Pydantic warns on ReadOnly TypedDict fields + tool_name: str # writable-ok: Pydantic warns on ReadOnly TypedDict fields + + +ToolMessageContentPart: TypeAlias = ( + ChatCompletionTextObject | ChatCompletionImageObject | ChatCompletionToolReferenceObject +) + + class ChatCompletionToolMessage(TypedDict): role: Literal["tool"] - content: str | Iterable[ChatCompletionTextObject | ChatCompletionImageObject] + content: str | Iterable[ToolMessageContentPart] # writable-ok: Pydantic warns on ReadOnly TypedDict fields tool_call_id: str diff --git a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py index 6ca48ce63b8..21b60d7a216 100644 --- a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py +++ b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py @@ -3903,3 +3903,39 @@ def test_stored_reasoning_items_win_over_thinking_blocks(): reasoning_items = [item for item in input_items if item.get("type") == "reasoning"] assert len(reasoning_items) == 1 assert reasoning_items[0]["id"] == "rs_real" + + +def test_convert_chat_completion_messages_to_responses_api_tool_result_with_tool_reference(): + """Tool-search tool_reference blocks have no Responses API equivalent: skip them, never stringify them.""" + from litellm.completion_extras.litellm_responses_transformation.transformation import ( + LiteLLMResponsesTransformationHandler, + ) + + handler = LiteLLMResponsesTransformationHandler() + + messages = [ + { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_abc123", + "type": "function", + "function": {"name": "ToolSearch", "arguments": '{"query": "web"}'}, + } + ], + }, + { + "role": "tool", + "tool_call_id": "call_abc123", + "content": [ + {"type": "tool_reference", "tool_name": "WebFetch"}, + {"type": "text", "text": "1 tool found"}, + ], + }, + ] + + response, _ = handler.convert_chat_completion_messages_to_responses_api(messages) + + function_call_output = next(item for item in response if item.get("type") == "function_call_output") + assert function_call_output["output"] == [{"type": "input_text", "text": "1 tool found"}] diff --git a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py index 6265779b90d..72d26f31c60 100644 --- a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py +++ b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py @@ -3578,3 +3578,52 @@ async def test_bedrock_converse_pdf_only_user_message_gets_text_block_async(): assert len(result) == 1 assert any("document" in block for block in result[0]["content"]) assert _text_blocks(result[0]) == [BEDROCK_DOCUMENT_PLACEHOLDER_TEXT] + + +def test_convert_to_anthropic_tool_result_keeps_tool_reference_blocks(): + from litellm.litellm_core_utils.prompt_templates.factory import convert_to_anthropic_tool_result + + result = convert_to_anthropic_tool_result( + { + "role": "tool", + "tool_call_id": "toolu_01", + "content": [ + {"type": "text", "text": "loaded"}, + {"type": "tool_reference", "tool_name": "WebFetch"}, + ], + } + ) + + assert result == { + "type": "tool_result", + "tool_use_id": "toolu_01", + "content": [ + {"type": "text", "text": "loaded"}, + {"type": "tool_reference", "tool_name": "WebFetch"}, + ], + } + + +def test_convert_gemini_tool_call_result_answers_tool_reference_only_result(): + """Every Gemini function call needs a function response, even when the tool result carries no text. + Fixes: https://github.com/BerriAI/litellm/issues/37462 + """ + result = convert_to_gemini_tool_call_result( + message=ChatCompletionToolMessage( + role="tool", + tool_call_id="toolu_01", + content=[{"type": "tool_reference", "tool_name": "WebFetch"}], + ), + last_message_with_tool_calls={ + "role": "assistant", + "tool_calls": [ + { + "id": "toolu_01", + "type": "function", + "function": {"name": "ToolSearch", "arguments": '{"query": "select:WebFetch"}'}, + } + ], + }, + ) + + assert result == {"function_response": {"name": "ToolSearch", "response": {"content": ""}}} diff --git a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py index 2b392456763..0f51b321e59 100644 --- a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py +++ b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py @@ -1818,3 +1818,72 @@ class TestAnthropicMessagesScanOnlyToolResults: assert guardrail.captured_inputs is not None assert guardrail.captured_inputs.get("images") == ["TOOL_IMG"] + + +class TestStructuredWriteBackKeepsToolResults: + """A guardrail rewrite must never leave a tool_use without its tool_result (Claude Code ToolSearch, LIT-6103).""" + + @staticmethod + def _claude_code_tool_search_turns(tool_result_content): + return [ + {"role": "user", "content": "load WebFetch for bob@example.com"}, + { + "role": "assistant", + "content": [ + { + "type": "tool_use", + "id": "toolu_01", + "name": "ToolSearch", + "input": {"query": "select:WebFetch"}, + } + ], + }, + { + "role": "user", + "content": [ + {"type": "tool_result", "tool_use_id": "toolu_01", "content": tool_result_content}, + {"type": "text", "text": "Now fetch the page."}, + ], + }, + ] + + @staticmethod + def _blocks(message): + return message["content"] if isinstance(message["content"], list) else [] + + @pytest.mark.parametrize( + ("tool_result_content", "expected_written_back_content"), + [ + ( + [{"type": "tool_reference", "tool_name": "WebFetch"}], + [{"type": "tool_reference", "tool_name": "WebFetch"}], + ), + ([], ""), + ], + ids=["tool_reference", "empty"], + ) + async def test_tool_result_stays_right_after_its_tool_use( + self, tool_result_content, expected_written_back_content + ): + handler = AnthropicMessagesHandler() + data = {"model": "claude-fable-5", "messages": self._claude_code_tool_search_turns(tool_result_content)} + + await handler.process_input_messages(data=data, guardrail_to_apply=MockStructuredMaskingGuardrail()) + + serialized = json.dumps(data["messages"]) + assert "bob@example.com" not in serialized + assert "" in serialized + + messages = data["messages"] + tool_use_index = next( + i for i, m in enumerate(messages) if any(b.get("type") == "tool_use" for b in self._blocks(m)) + ) + answer = messages[tool_use_index + 1] + assert answer["role"] == "user" + assert answer["content"][0] == { + "type": "tool_result", + "tool_use_id": "toolu_01", + "content": expected_written_back_content, + } + later_blocks = [b for m in messages[tool_use_index + 1 :] for b in self._blocks(m)] + assert {"type": "text", "text": "Now fetch the page."} in later_blocks diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py index ee09baf28b6..9c88d7a67bb 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py @@ -3997,3 +3997,72 @@ def test_translate_anthropic_messages_to_openai_carries_midturn_system_prompt_ca assert result == [ {"role": "system", "content": [{"type": "text", "text": "fix", "prompt_cache_breakpoint": explicit}]} ] + + +def _tool_reference_block(tool_name="WebFetch"): + return {"type": "tool_reference", "tool_name": tool_name} + + +def test_tool_result_tool_reference_is_carried_through_untouched(): + adapter = LiteLLMAnthropicMessagesAdapter() + + result = adapter.translate_anthropic_messages_to_openai( + messages=[ + _anthropic_tool_use_turn("toolu_01"), + _anthropic_tool_result_turn({"toolu_01": [_tool_reference_block()]}), + ] + ) + + assert [m["role"] for m in result] == ["assistant", "tool"] + assert result[1]["tool_call_id"] == "toolu_01" + assert result[1]["content"] == [{"type": "tool_reference", "tool_name": "WebFetch"}] + + +def test_tool_result_text_beside_tool_reference_keeps_both_parts_in_order(): + adapter = LiteLLMAnthropicMessagesAdapter() + + result = adapter.translate_anthropic_messages_to_openai( + messages=[ + _anthropic_tool_use_turn("toolu_01"), + _anthropic_tool_result_turn( + {"toolu_01": [{"type": "text", "text": "loaded"}, _tool_reference_block("Grep")]} + ), + ] + ) + + assert result[1]["content"] == [ + {"type": "text", "text": "loaded"}, + {"type": "tool_reference", "tool_name": "Grep"}, + ] + + +@pytest.mark.parametrize( + "tool_result_content", + [ + [], + None, + "", + {"not": "a list"}, + [{"type": "future_block", "payload": 1}], + [{"type": "search_result", "source": "https://example.com", "title": "t", "content": []}], + ], + ids=["empty_list", "null", "empty_string", "non_list", "unknown_block", "search_result_only"], +) +def test_tool_result_without_translatable_content_still_answers_its_tool_use(tool_result_content): + adapter = LiteLLMAnthropicMessagesAdapter() + + result = adapter.translate_anthropic_messages_to_openai( + messages=[ + _anthropic_tool_use_turn("toolu_01"), + { + "role": "user", + "content": [{"type": "tool_result", "tool_use_id": "toolu_01", "content": tool_result_content}], + }, + ] + ) + + assert result == [ + result[0], + {"role": "tool", "tool_call_id": "toolu_01", "content": ""}, + ] + assert result[0]["role"] == "assistant" diff --git a/type-discipline-budget.json b/type-discipline-budget.json index 6fd7828906c..94f671d7c3c 100644 --- a/type-discipline-budget.json +++ b/type-discipline-budget.json @@ -3,7 +3,7 @@ "limit": 22733 }, "LIT002": { - "limit": 26863 + "limit": 26860 }, "LIT003": { "limit": 269 @@ -33,6 +33,6 @@ "limit": 5583 }, "LIT012": { - "limit": 4510 + "limit": 4509 } } diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 4ccf7b59fbc..aafbb2eb6b2 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -23643,7 +23643,7 @@ export interface components { /** ChatCompletionToolMessage */ ChatCompletionToolMessage: { /** Content */ - content: string | (components["schemas"]["ChatCompletionTextObject"] | components["schemas"]["ChatCompletionImageObject"])[]; + content: string | (components["schemas"]["ChatCompletionTextObject"] | components["schemas"]["ChatCompletionImageObject"] | components["schemas"]["ChatCompletionToolReferenceObject"])[]; /** * Role * @constant @@ -23672,6 +23672,19 @@ export interface components { /** Strict */ strict?: boolean; }; + /** + * ChatCompletionToolReferenceObject + * @description Anthropic tool-search result block, carried through untouched so it survives a round trip. + */ + ChatCompletionToolReferenceObject: { + /** Tool Name */ + tool_name: string; + /** + * Type + * @constant + */ + type: "tool_reference"; + }; /** ChatCompletionUserMessage */ ChatCompletionUserMessage: { cache_control?: components["schemas"]["ChatCompletionCachedContent"]; From edde2e50efe848f0dfa600af5a5d936e6443e2c7 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 23:43:41 -0700 Subject: [PATCH 28/44] fix(openai): drop tool_reference parts from tool messages at the chat boundary OpenAI's chat completions API rejects tool_reference content parts in role tool messages, so a mixed text plus reference tool result carried through the Anthropic adapter turned a previously working request into a 400 on chat-routed OpenAI and Azure deployments. Strip the reference parts there, keeping a reference-only result as an empty-text tool message so the preceding tool_call stays answered, mirroring the Responses bridge skip. --- .../prompt_templates/common_utils.py | 40 +++++++++++ litellm/llms/azure/chat/gpt_transformation.py | 4 +- .../llms/openai/chat/gpt_transformation.py | 4 +- ...ore_utils_prompt_templates_common_utils.py | 67 +++++++++++++++++++ .../test_azure_chat_gpt_transformation.py | 29 ++++++++ .../chat/test_openai_gpt_transformation.py | 63 +++++++++++++++++ 6 files changed, 205 insertions(+), 2 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 748347fe938..f0e5086b660 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -1747,6 +1747,46 @@ def hoist_images_from_tool_messages( ] +def _is_tool_reference_part(part: object) -> bool: + return isinstance(part, dict) and part.get("type") == "tool_reference" + + +def _tool_message_carries_tool_reference(message: AllMessageValues) -> bool: + if message.get("role") != "tool": + return False + content = message.get("content") + return isinstance(content, list) and any(_is_tool_reference_part(part) for part in content) + + +def _drop_tool_reference_parts(message: AllMessageValues) -> AllMessageValues: + if not _tool_message_carries_tool_reference(message): + return message + content = cast(list, message.get("content")) # cast-ok: shape checked by _tool_message_carries_tool_reference + remaining_parts = [ # mutable-ok: tool message content must stay a json list + part for part in content if not _is_tool_reference_part(part) + ] + new_content = remaining_parts if remaining_parts else "" + rewritten = {**message, "content": new_content} # mutable-ok: chat messages are plain json dicts + return cast(AllMessageValues, rewritten) # cast-ok: dict spread keeps keys like cache_control + + +def drop_tool_reference_parts_from_tool_messages( + messages: list[AllMessageValues], # mutable-ok: message pipelines type messages as mutable lists +) -> list[AllMessageValues]: # mutable-ok: message pipelines type messages as mutable lists + """ + Remove tool_reference content parts from role:"tool" messages. + + The OpenAI chat spec only accepts text in tool messages, so a tool_reference + part carried through the Anthropic adapter makes strict providers reject the + request. The reference names an already-declared tool rather than carrying + content, so it is dropped; a reference-only result keeps its tool message with + empty text so the preceding tool_call stays answered. + """ + if not any(_tool_message_carries_tool_reference(message) for message in messages): + return messages + return [_drop_tool_reference_parts(message) for message in messages] # mutable-ok: pipelines mutate message lists + + def _attempt_json_repair(s: str) -> Any | None: """ Attempt to repair truncated JSON produced by LLM tool calls. diff --git a/litellm/llms/azure/chat/gpt_transformation.py b/litellm/llms/azure/chat/gpt_transformation.py index 0d50609555a..30fc3635d9d 100644 --- a/litellm/llms/azure/chat/gpt_transformation.py +++ b/litellm/llms/azure/chat/gpt_transformation.py @@ -4,6 +4,7 @@ from httpx._models import Headers, Response import litellm from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, hoist_images_from_tool_messages, ) from litellm.litellm_core_utils.prompt_templates.factory import ( @@ -252,7 +253,8 @@ class AzureOpenAIConfig(BaseConfig): litellm_params: dict, headers: dict, ) -> dict: - azure_messages: Final = convert_to_azure_openai_messages(hoist_images_from_tool_messages(messages)) + stripped_messages: Final = drop_tool_reference_parts_from_tool_messages(messages) + azure_messages: Final = convert_to_azure_openai_messages(hoist_images_from_tool_messages(stripped_messages)) return { "model": model, "messages": azure_messages, diff --git a/litellm/llms/openai/chat/gpt_transformation.py b/litellm/llms/openai/chat/gpt_transformation.py index 16fd042cb2f..9b7c5a3f857 100644 --- a/litellm/llms/openai/chat/gpt_transformation.py +++ b/litellm/llms/openai/chat/gpt_transformation.py @@ -18,6 +18,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo _should_convert_tool_call_to_json_mode, ) from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, get_tool_call_names, hoist_images_from_tool_messages, ) @@ -336,7 +337,8 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig): self, messages: list[AllMessageValues], model: str, is_async: bool = False ) -> list[AllMessageValues] | Coroutine[Any, Any, list[AllMessageValues]]: """OpenAI no longer supports image_url as a string, so we need to convert it to a dict""" - hoisted_messages: Final = hoist_images_from_tool_messages(messages) + stripped_messages: Final = drop_tool_reference_parts_from_tool_messages(messages) + hoisted_messages: Final = hoist_images_from_tool_messages(stripped_messages) async def _async_transform(): for message in hoisted_messages: diff --git a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py index 44f91c98d81..aec6d12069f 100644 --- a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py +++ b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py @@ -1027,3 +1027,70 @@ def test_update_messages_xlitellm_decode_does_not_override_mapping(): updated = update_messages_with_model_file_ids(messages, "model-A", mapping) assert updated[0]["content"][0]["file"]["file_id"] == "provider-explicit-id" + + +def test_drop_tool_reference_parts_keeps_text_parts(): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, + ) + + messages = [ + _assistant_tool_call_msg("call_1"), + _tool_msg( + [ + {"type": "text", "text": "WebFetch tool loaded successfully."}, + {"type": "tool_reference", "tool_name": "WebFetch"}, + ] + ), + ] + + result = drop_tool_reference_parts_from_tool_messages(messages) + + assert result[1]["content"] == [{"type": "text", "text": "WebFetch tool loaded successfully."}] + assert result[1]["tool_call_id"] == "call_1" + + +def test_drop_tool_reference_parts_reference_only_becomes_empty_text(): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, + ) + + messages = [ + _assistant_tool_call_msg("call_1"), + _tool_msg([{"type": "tool_reference", "tool_name": "WebFetch"}]), + ] + + result = drop_tool_reference_parts_from_tool_messages(messages) + + assert result[1] == {"role": "tool", "tool_call_id": "call_1", "content": ""} + + +def test_drop_tool_reference_parts_without_references_passes_through(): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, + ) + + messages = [ + _assistant_tool_call_msg("call_1"), + _tool_msg([{"type": "text", "text": "plain result"}]), + ] + + assert drop_tool_reference_parts_from_tool_messages(messages) is messages + + +def test_drop_tool_reference_parts_leaves_non_tool_messages_alone(): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_tool_reference_parts_from_tool_messages, + ) + + user_message = {"role": "user", "content": [{"type": "tool_reference", "tool_name": "WebFetch"}]} + messages = [ + user_message, + _assistant_tool_call_msg("call_1"), + _tool_msg([{"type": "tool_reference", "tool_name": "WebFetch"}]), + ] + + result = drop_tool_reference_parts_from_tool_messages(messages) + + assert result[0] == user_message + assert result[2]["content"] == "" diff --git a/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py b/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py index ad34199c4c6..2cf7cd142d6 100644 --- a/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py +++ b/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py @@ -102,6 +102,35 @@ def test_transform_request_hoists_tool_message_image(): ] +def test_transform_request_drops_tool_reference_parts(): + """Azure's transform_request shares the tool-message sanitizing with OpenAI: + tool_reference parts are dropped, a reference-only result keeps its tool + message with empty text (#37462 round trip).""" + messages = [ + {"role": "user", "content": "load the WebFetch tool"}, + { + "role": "assistant", + "content": None, + "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "ToolSearch", "arguments": "{}"}}], + }, + { + "role": "tool", + "tool_call_id": "call_1", + "content": [{"type": "tool_reference", "tool_name": "WebFetch"}], + }, + ] + + request = AzureOpenAIConfig().transform_request( + model="gpt-4o", + messages=messages, + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][2]["content"] == "" + + @pytest.mark.parametrize( "model, emitted_key, absent_key", [ diff --git a/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py b/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py index f4c38f8f797..3f346b5e8e7 100644 --- a/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py +++ b/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py @@ -869,6 +869,69 @@ class TestToolMessageImageHoisting: assert result[3]["content"] == self.HOISTED_USER_CONTENT +class TestToolReferenceStripping: + """transform_request drops tool_reference parts from tool messages: OpenAI's + chat API rejects them, and the reference names an already-declared tool + rather than carrying content (#37462 round trip).""" + + def setup_method(self): + self.config = OpenAIGPTConfig() + + def _messages_with_tool_reference(self, extra_parts=()): + return [ + {"role": "user", "content": "load the WebFetch tool"}, + { + "role": "assistant", + "content": None, + "tool_calls": [ + {"id": "call_1", "type": "function", "function": {"name": "ToolSearch", "arguments": "{}"}} + ], + }, + { + "role": "tool", + "tool_call_id": "call_1", + "content": [*extra_parts, {"type": "tool_reference", "tool_name": "WebFetch"}], + }, + ] + + def test_transform_request_keeps_text_and_drops_reference(self): + request = self.config.transform_request( + model="gpt-4.1", + messages=self._messages_with_tool_reference(extra_parts=({"type": "text", "text": "loaded"},)), + optional_params={}, + litellm_params={}, + headers={}, + ) + + tool_message = request["messages"][2] + assert tool_message["content"] == [{"type": "text", "text": "loaded"}] + assert tool_message["tool_call_id"] == "call_1" + + def test_transform_request_reference_only_keeps_tool_message_with_empty_text(self): + request = self.config.transform_request( + model="gpt-4.1", + messages=self._messages_with_tool_reference(), + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert [m.get("role") for m in request["messages"]] == ["user", "assistant", "tool"] + assert request["messages"][2]["content"] == "" + + @pytest.mark.asyncio + async def test_async_transform_request_drops_reference(self): + request = await self.config.async_transform_request( + model="gpt-4.1", + messages=self._messages_with_tool_reference(), + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][2]["content"] == "" + + class TestOpenAIPromptCacheBreakpointChatPath: """Chat-path shape for OpenAI explicit prompt caching (#37509).""" From b4c6e01fcc83cc8904517d4485fef1eccc9f66b6 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 01:12:12 -0700 Subject: [PATCH 29/44] feat(together_ai): add zai-org/GLM-5.3-Flash to the model registry Adds pricing (0.15/0.50 per 1M tokens, 0.03 cached read), the 1M context window, and capability flags (tools, parallel tools, tool choice, response schema, reasoning, vision) for Together AI's zai-org/GLM-5.3-Flash, mirrored into the backup cost map, with exact-value regression tests. --- .../model_prices_and_context_window_backup.json | 17 +++++++++++++++++ model_prices_and_context_window.json | 17 +++++++++++++++++ .../test_together_ai_model_metadata.py | 16 ++++++++++++++++ 3 files changed, 50 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index dd367e875de..7f8c464f5b9 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -38770,6 +38770,23 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "together_ai/zai-org/GLM-5.3-Flash": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "together_ai", + "max_input_tokens": 1048575, + "max_output_tokens": 1048575, + "max_tokens": 1048575, + "mode": "chat", + "output_cost_per_token": 5e-07, + "source": "https://docs.together.ai/docs/serverless-models", + "supports_function_calling": true, + "supports_parallel_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "tts-1": { "input_cost_per_character": 1.5e-05, "litellm_provider": "openai", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index dd367e875de..7f8c464f5b9 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -38770,6 +38770,23 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "together_ai/zai-org/GLM-5.3-Flash": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "together_ai", + "max_input_tokens": 1048575, + "max_output_tokens": 1048575, + "max_tokens": 1048575, + "mode": "chat", + "output_cost_per_token": 5e-07, + "source": "https://docs.together.ai/docs/serverless-models", + "supports_function_calling": true, + "supports_parallel_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "tts-1": { "input_cost_per_character": 1.5e-05, "litellm_provider": "openai", diff --git a/tests/test_litellm/test_together_ai_model_metadata.py b/tests/test_litellm/test_together_ai_model_metadata.py index 5a0aadf4737..fbeb0c35fca 100644 --- a/tests/test_litellm/test_together_ai_model_metadata.py +++ b/tests/test_litellm/test_together_ai_model_metadata.py @@ -15,6 +15,7 @@ COST_MAP_ADAPTER: Final = TypeAdapter(CostMap) SERVERLESS_CHAT_MODELS: Final = ( "together_ai/moonshotai/Kimi-K3", "together_ai/zai-org/GLM-5.2", + "together_ai/zai-org/GLM-5.3-Flash", "together_ai/deepseek-ai/DeepSeek-V4-Pro", "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813", "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731", @@ -110,6 +111,21 @@ def test_together_glm_52_pricing(cost_map: CostMap): assert info["supports_reasoning"] is True +def test_together_glm_53_flash_pricing_and_capabilities(cost_map: CostMap): + info = cost_map["together_ai/zai-org/GLM-5.3-Flash"] + assert info["input_cost_per_token"] == 1.5e-07 + assert info["output_cost_per_token"] == 5e-07 + assert info["cache_read_input_token_cost"] == 3e-08 + assert info["max_input_tokens"] == 1048575 + assert info["max_output_tokens"] == 1048575 + assert info["supports_function_calling"] is True + assert info["supports_parallel_function_calling"] is True + assert info["supports_tool_choice"] is True + assert info["supports_response_schema"] is True + assert info["supports_vision"] is True + assert info["supports_reasoning"] is True + + def test_together_multilingual_e5_embedding_entry(cost_map: CostMap): info = cost_map["together_ai/intfloat/multilingual-e5-large-instruct"] assert info["mode"] == "embedding" From bcb6a0a998c5059b6668c9d122bdd1ed33ab8a38 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 01:27:32 -0700 Subject: [PATCH 30/44] test(together_ai): assert fail-open supported params for models missing from the registry --- tests/llm_translation/test_together_ai.py | 20 ++++++-------------- 1 file changed, 6 insertions(+), 14 deletions(-) diff --git a/tests/llm_translation/test_together_ai.py b/tests/llm_translation/test_together_ai.py index c371caefa5e..fd7ad40ed11 100644 --- a/tests/llm_translation/test_together_ai.py +++ b/tests/llm_translation/test_together_ai.py @@ -23,26 +23,18 @@ class TestTogetherAI(BaseLLMChatTest): pass @pytest.mark.parametrize( - "model, expected_bool", + "model", [ - ("meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo", True), - ("nvidia/Llama-3.1-Nemotron-70B-Instruct-HF", False), + "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo", + "nvidia/Llama-3.1-Nemotron-70B-Instruct-HF", ], ) - def test_get_supported_response_format_together_ai( - self, model: str, expected_bool: bool - ) -> None: + def test_get_supported_response_format_together_ai(self, model: str) -> None: os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True" litellm.model_cost = litellm.get_model_cost_map(url="") optional_params = litellm.get_supported_openai_params( model, custom_llm_provider="together_ai" ) - # Mapped provider assert isinstance(optional_params, list) - - if expected_bool: - assert "response_format" in optional_params - assert "tools" in optional_params - else: - assert "response_format" not in optional_params - assert "tools" not in optional_params + assert "response_format" in optional_params + assert "tools" in optional_params From 5d26ae0fcd77f50ec5e210b8b529dc77370da594 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 27 Aug 2026 13:16:12 +0000 Subject: [PATCH 31/44] fix(model_prices): absorb Databricks/Z.AI and xAI registry PRs, add Together and Azure deprecation dates Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- ...odel_prices_and_context_window_backup.json | 203 ++++++++---------- model_prices_and_context_window.json | 203 ++++++++---------- .../test_databricks_cost_calculator.py | 2 + .../llms/xai/test_xai_model_registry.py | 75 +++++++ .../test_together_ai_model_metadata.py | 8 +- 5 files changed, 261 insertions(+), 230 deletions(-) create mode 100644 tests/test_litellm/llms/xai/test_xai_model_registry.py diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index cfa06ff5f81..0d867955544 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -15167,6 +15167,34 @@ "output_dbu_cost_per_token": 7.143e-06, "source": "https://www.databricks.com/product/pricing/foundation-model-serving" }, + "databricks/databricks-glm-5-2": { + "cache_creation_input_token_cost": 1.4e-06, + "cache_read_input_token_cost": 2.5998e-07, + "input_cost_per_token": 1.4e-06, + "input_dbu_cost_per_token": 2e-05, + "litellm_provider": "databricks", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "metadata": { + "notes": "Input/output cost per token is dbu cost * $0.070. Billing reads the per-token dollar fields; the '*_dbu_cost_per_token' fields are the published Databricks rates, kept for reference." + }, + "mode": "chat", + "output_cost_per_token": 4.39999e-06, + "output_dbu_cost_per_token": 6.2857e-05, + "source": "https://www.databricks.com/product/pricing/foundation-model-serving", + "supported_modalities": [ + "text" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, "databricks/databricks-gpt-5": { "cache_creation_input_token_cost": 1.24999e-06, "cache_read_input_token_cost": 1.2502e-07, @@ -15434,6 +15462,35 @@ "output_vector_size": 1024, "source": "https://www.databricks.com/product/pricing/foundation-model-serving" }, + "databricks/databricks-kimi-k3": { + "cache_creation_input_token_cost": 2.99999e-06, + "cache_read_input_token_cost": 3.0002e-07, + "input_cost_per_token": 2.99999e-06, + "input_dbu_cost_per_token": 4.2857e-05, + "litellm_provider": "databricks", + "max_input_tokens": 1000000, + "max_output_tokens": 1048576, + "max_tokens": 1048576, + "metadata": { + "notes": "Input/output cost per token is dbu cost * $0.070. Billing reads the per-token dollar fields; the '*_dbu_cost_per_token' fields are the published Databricks rates, kept for reference." + }, + "mode": "chat", + "output_cost_per_token": 1.500002e-05, + "output_dbu_cost_per_token": 0.000214286, + "source": "https://www.databricks.com/product/pricing/foundation-model-serving", + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": true + }, "databricks/databricks-llama-2-70b-chat": { "cache_creation_input_token_cost": 5.0001e-07, "cache_read_input_token_cost": 5.0001e-07, @@ -38329,7 +38386,7 @@ "max_output_tokens": 20480, "max_tokens": 20480, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 7e-06, @@ -38358,7 +38415,7 @@ "max_output_tokens": 8192, "max_tokens": 8192, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 1.25e-06, @@ -38373,7 +38430,7 @@ "litellm_provider": "together_ai", "max_tokens": 16384, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 1.7e-06, @@ -38794,6 +38851,7 @@ "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", @@ -38886,6 +38944,7 @@ "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", @@ -38921,6 +38980,7 @@ "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", @@ -38938,6 +38998,7 @@ "supports_tool_choice": true }, "together_ai/pearl-ai/gemma-4-31b-it": { + "deprecation_date": "2026-08-27", "input_cost_per_token": 2.8e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -43788,85 +43849,6 @@ "/v1/audio/transcriptions" ] }, - "xai/grok-2": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-1212": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-latest": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-vision": { - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "xai/grok-2-vision-1212": { - "deprecation_date": "2026-02-28", - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "xai/grok-2-vision-latest": { - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "xai/grok-3": { "cache_read_input_token_cost": 7.5e-07, "input_cost_per_token": 3e-06, @@ -44252,7 +44234,7 @@ "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, - "mode": "chat", + "mode": "responses", "output_cost_per_token": 2.5e-06, "source": "https://docs.x.ai/docs/models", "supports_function_calling": true, @@ -44264,7 +44246,10 @@ "input_cost_per_token_above_200k_tokens": 2.5e-06, "output_cost_per_token_above_200k_tokens": 5e-06, "cache_read_input_token_cost_above_200k_tokens": 4e-07, - "supports_response_schema": true + "supports_response_schema": true, + "supported_endpoints": [ + "/v1/responses" + ] }, "xai/grok-4.20-beta-0309-reasoning": { "cache_read_input_token_cost": 2e-07, @@ -44433,19 +44418,6 @@ "supports_vision": true, "supports_web_search": true }, - "xai/grok-beta": { - "input_cost_per_token": 5e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "xai/grok-code-fast": { "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1e-06, @@ -44509,20 +44481,6 @@ "supports_vision": true, "deprecation_date": "2026-05-15" }, - "xai/grok-vision-beta": { - "input_cost_per_image": 5e-06, - "input_cost_per_token": 5e-06, - "litellm_provider": "xai", - "max_input_tokens": 8192, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "zai.glm-4.7": { "input_cost_per_token": 6e-07, "litellm_provider": "bedrock_converse", @@ -44581,6 +44539,21 @@ "supports_tool_choice": true, "source": "https://docs.z.ai/guides/overview/pricing" }, + "zai/glm-5.3": { + "cache_creation_input_token_cost": 0, + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "zai", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.z.ai/guides/overview/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true + }, "zai/glm-5.1": { "cache_creation_input_token_cost": 0, "cache_read_input_token_cost": 2.6e-07, @@ -44786,6 +44759,7 @@ ] }, "azure/sora-2": { + "deprecation_date": "2026-10-15", "litellm_provider": "azure", "mode": "video_generation", "output_cost_per_video_per_second": 0.1, @@ -51240,7 +51214,7 @@ "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, - "mode": "chat", + "mode": "responses", "output_cost_per_token": 2.5e-06, "source": "https://docs.x.ai/docs/models", "supports_function_calling": true, @@ -51252,7 +51226,10 @@ "input_cost_per_token_above_200k_tokens": 2.5e-06, "output_cost_per_token_above_200k_tokens": 5e-06, "cache_read_input_token_cost_above_200k_tokens": 4e-07, - "supports_response_schema": true + "supports_response_schema": true, + "supported_endpoints": [ + "/v1/responses" + ] }, "xai/grok-build-0.1": { "cache_read_input_token_cost": 2e-07, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index cfa06ff5f81..0d867955544 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -15167,6 +15167,34 @@ "output_dbu_cost_per_token": 7.143e-06, "source": "https://www.databricks.com/product/pricing/foundation-model-serving" }, + "databricks/databricks-glm-5-2": { + "cache_creation_input_token_cost": 1.4e-06, + "cache_read_input_token_cost": 2.5998e-07, + "input_cost_per_token": 1.4e-06, + "input_dbu_cost_per_token": 2e-05, + "litellm_provider": "databricks", + "max_input_tokens": 1000000, + "max_output_tokens": 131072, + "max_tokens": 131072, + "metadata": { + "notes": "Input/output cost per token is dbu cost * $0.070. Billing reads the per-token dollar fields; the '*_dbu_cost_per_token' fields are the published Databricks rates, kept for reference." + }, + "mode": "chat", + "output_cost_per_token": 4.39999e-06, + "output_dbu_cost_per_token": 6.2857e-05, + "source": "https://www.databricks.com/product/pricing/foundation-model-serving", + "supported_modalities": [ + "text" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": false + }, "databricks/databricks-gpt-5": { "cache_creation_input_token_cost": 1.24999e-06, "cache_read_input_token_cost": 1.2502e-07, @@ -15434,6 +15462,35 @@ "output_vector_size": 1024, "source": "https://www.databricks.com/product/pricing/foundation-model-serving" }, + "databricks/databricks-kimi-k3": { + "cache_creation_input_token_cost": 2.99999e-06, + "cache_read_input_token_cost": 3.0002e-07, + "input_cost_per_token": 2.99999e-06, + "input_dbu_cost_per_token": 4.2857e-05, + "litellm_provider": "databricks", + "max_input_tokens": 1000000, + "max_output_tokens": 1048576, + "max_tokens": 1048576, + "metadata": { + "notes": "Input/output cost per token is dbu cost * $0.070. Billing reads the per-token dollar fields; the '*_dbu_cost_per_token' fields are the published Databricks rates, kept for reference." + }, + "mode": "chat", + "output_cost_per_token": 1.500002e-05, + "output_dbu_cost_per_token": 0.000214286, + "source": "https://www.databricks.com/product/pricing/foundation-model-serving", + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": true + }, "databricks/databricks-llama-2-70b-chat": { "cache_creation_input_token_cost": 5.0001e-07, "cache_read_input_token_cost": 5.0001e-07, @@ -38329,7 +38386,7 @@ "max_output_tokens": 20480, "max_tokens": 20480, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 7e-06, @@ -38358,7 +38415,7 @@ "max_output_tokens": 8192, "max_tokens": 8192, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 1.25e-06, @@ -38373,7 +38430,7 @@ "litellm_provider": "together_ai", "max_tokens": 16384, "metadata": { - "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro" + "successor": "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813" }, "mode": "chat", "output_cost_per_token": 1.7e-06, @@ -38794,6 +38851,7 @@ "supports_tool_choice": true }, "together_ai/deepseek-ai/DeepSeek-V4-Pro": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1.74e-06, "litellm_provider": "together_ai", @@ -38886,6 +38944,7 @@ "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 1.9e-07, "input_cost_per_token": 9.5e-07, "litellm_provider": "together_ai", @@ -38921,6 +38980,7 @@ "supports_vision": true }, "together_ai/nvidia/nemotron-3-ultra-550b-a55b": { + "deprecation_date": "2026-08-27", "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 6e-07, "litellm_provider": "together_ai", @@ -38938,6 +38998,7 @@ "supports_tool_choice": true }, "together_ai/pearl-ai/gemma-4-31b-it": { + "deprecation_date": "2026-08-27", "input_cost_per_token": 2.8e-07, "litellm_provider": "together_ai", "max_input_tokens": 262144, @@ -43788,85 +43849,6 @@ "/v1/audio/transcriptions" ] }, - "xai/grok-2": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-1212": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-latest": { - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_web_search": true - }, - "xai/grok-2-vision": { - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "xai/grok-2-vision-1212": { - "deprecation_date": "2026-02-28", - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "xai/grok-2-vision-latest": { - "input_cost_per_image": 2e-06, - "input_cost_per_token": 2e-06, - "litellm_provider": "xai", - "max_input_tokens": 32768, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 1e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "xai/grok-3": { "cache_read_input_token_cost": 7.5e-07, "input_cost_per_token": 3e-06, @@ -44252,7 +44234,7 @@ "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, - "mode": "chat", + "mode": "responses", "output_cost_per_token": 2.5e-06, "source": "https://docs.x.ai/docs/models", "supports_function_calling": true, @@ -44264,7 +44246,10 @@ "input_cost_per_token_above_200k_tokens": 2.5e-06, "output_cost_per_token_above_200k_tokens": 5e-06, "cache_read_input_token_cost_above_200k_tokens": 4e-07, - "supports_response_schema": true + "supports_response_schema": true, + "supported_endpoints": [ + "/v1/responses" + ] }, "xai/grok-4.20-beta-0309-reasoning": { "cache_read_input_token_cost": 2e-07, @@ -44433,19 +44418,6 @@ "supports_vision": true, "supports_web_search": true }, - "xai/grok-beta": { - "input_cost_per_token": 5e-06, - "litellm_provider": "xai", - "max_input_tokens": 131072, - "max_output_tokens": 131072, - "max_tokens": 131072, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "xai/grok-code-fast": { "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 1e-06, @@ -44509,20 +44481,6 @@ "supports_vision": true, "deprecation_date": "2026-05-15" }, - "xai/grok-vision-beta": { - "input_cost_per_image": 5e-06, - "input_cost_per_token": 5e-06, - "litellm_provider": "xai", - "max_input_tokens": 8192, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supports_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "zai.glm-4.7": { "input_cost_per_token": 6e-07, "litellm_provider": "bedrock_converse", @@ -44581,6 +44539,21 @@ "supports_tool_choice": true, "source": "https://docs.z.ai/guides/overview/pricing" }, + "zai/glm-5.3": { + "cache_creation_input_token_cost": 0, + "cache_read_input_token_cost": 2.6e-07, + "input_cost_per_token": 1.4e-06, + "litellm_provider": "zai", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.4e-06, + "source": "https://docs.z.ai/guides/overview/pricing", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true + }, "zai/glm-5.1": { "cache_creation_input_token_cost": 0, "cache_read_input_token_cost": 2.6e-07, @@ -44786,6 +44759,7 @@ ] }, "azure/sora-2": { + "deprecation_date": "2026-10-15", "litellm_provider": "azure", "mode": "video_generation", "output_cost_per_video_per_second": 0.1, @@ -51240,7 +51214,7 @@ "max_input_tokens": 1000000, "max_output_tokens": 1000000, "max_tokens": 1000000, - "mode": "chat", + "mode": "responses", "output_cost_per_token": 2.5e-06, "source": "https://docs.x.ai/docs/models", "supports_function_calling": true, @@ -51252,7 +51226,10 @@ "input_cost_per_token_above_200k_tokens": 2.5e-06, "output_cost_per_token_above_200k_tokens": 5e-06, "cache_read_input_token_cost_above_200k_tokens": 4e-07, - "supports_response_schema": true + "supports_response_schema": true, + "supported_endpoints": [ + "/v1/responses" + ] }, "xai/grok-build-0.1": { "cache_read_input_token_cost": 2e-07, diff --git a/tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py b/tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py index 21f047b753c..29ad8ee4b6e 100644 --- a/tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py +++ b/tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py @@ -61,6 +61,8 @@ PUBLISHED_DBU_PER_MILLION: Final = { "databricks/databricks-gemini-3-1-flash-lite": ("4.464", "26.786", "4.464", "0.446"), "databricks/databricks-gemini-2-5-pro": ("22.321", "178.571", "22.321", "2.232"), "databricks/databricks-gemini-2-5-flash": ("5.357", "44.643", "5.357", "0.536"), + "databricks/databricks-kimi-k3": ("42.857", "214.286", "42.857", "4.286"), + "databricks/databricks-glm-5-2": ("20.000", "62.857", "20.000", "3.714"), } PROMOTIONAL_DISCOUNT: Final = 0.80 PROMOTION_EXPIRES: Final = "2027-01-31" diff --git a/tests/test_litellm/llms/xai/test_xai_model_registry.py b/tests/test_litellm/llms/xai/test_xai_model_registry.py new file mode 100644 index 00000000000..25b2002968d --- /dev/null +++ b/tests/test_litellm/llms/xai/test_xai_model_registry.py @@ -0,0 +1,75 @@ +""" +Registry regression tests for xAI entries in the model cost map. +""" + +from __future__ import annotations + +import json +from pathlib import Path + +import pytest + +REPO_ROOT = Path(__file__).parents[4] +PRICES_PATH = REPO_ROOT / "model_prices_and_context_window.json" +BACKUP_PRICES_PATH = REPO_ROOT / "litellm" / "model_prices_and_context_window_backup.json" + +# Retired by xAI and no longer served: requests to these slugs 404 rather than +# redirecting, and they are absent from https://docs.x.ai/docs/models +RETIRED_MODELS = ( + "xai/grok-2", + "xai/grok-2-1212", + "xai/grok-2-latest", + "xai/grok-2-vision", + "xai/grok-2-vision-1212", + "xai/grok-2-vision-latest", + "xai/grok-beta", + "xai/grok-vision-beta", +) + +# https://docs.x.ai/developers/model-capabilities/text/multi-agent +# "The multi-agent model does not work with the OpenAI Chat Completions API." +RESPONSES_ONLY_MODELS = ( + "xai/grok-4.20-multi-agent-0309", + "xai/grok-4.20-multi-agent-beta-0309", +) + +MAP_PATHS = (PRICES_PATH, BACKUP_PRICES_PATH) + + +@pytest.fixture(scope="module", params=[p.name for p in MAP_PATHS]) +def cost_map(request: pytest.FixtureRequest) -> dict: + path = next(p for p in MAP_PATHS if p.name == request.param) + return json.loads(path.read_text(encoding="utf-8")) + + +@pytest.mark.parametrize("model", RETIRED_MODELS) +def test_retired_xai_models_are_not_advertised(cost_map: dict, model: str): + assert model not in cost_map + + +@pytest.mark.parametrize("model", RESPONSES_ONLY_MODELS) +def test_multi_agent_models_are_responses_only(cost_map: dict, model: str): + entry = cost_map[model] + assert entry["supported_endpoints"] == ["/v1/responses"] + assert entry["mode"] == "responses" + assert "/v1/chat/completions" not in entry["supported_endpoints"] + + +def test_surviving_xai_chat_models_still_serve_chat_completions(cost_map: dict): + """Guard against the removal above over-reaching into live models.""" + chat_models = [ + key + for key, value in cost_map.items() + if isinstance(value, dict) and value.get("litellm_provider") == "xai" and value.get("mode") == "chat" + ] + assert "xai/grok-4.3" in chat_models + assert "xai/grok-4.6" in chat_models + assert not any(key.startswith("xai/grok-2") for key in chat_models) + + +def test_both_cost_maps_agree_on_xai_entries(): + prices = json.loads(PRICES_PATH.read_text(encoding="utf-8")) + backup = json.loads(BACKUP_PRICES_PATH.read_text(encoding="utf-8")) + xai_keys = {k for k, v in prices.items() if isinstance(v, dict) and v.get("litellm_provider") == "xai"} + assert xai_keys + assert {k: prices[k] for k in xai_keys} == {k: backup[k] for k in xai_keys} diff --git a/tests/test_litellm/test_together_ai_model_metadata.py b/tests/test_litellm/test_together_ai_model_metadata.py index 60e4b8ddb4d..32d0769becc 100644 --- a/tests/test_litellm/test_together_ai_model_metadata.py +++ b/tests/test_litellm/test_together_ai_model_metadata.py @@ -15,10 +15,8 @@ COST_MAP_ADAPTER: Final = TypeAdapter(CostMap) SERVERLESS_CHAT_MODELS: Final = ( "together_ai/moonshotai/Kimi-K3", "together_ai/zai-org/GLM-5.2", - "together_ai/deepseek-ai/DeepSeek-V4-Pro", "together_ai/deepseek-ai/DeepSeek-V4-Pro-0813", "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731", - "together_ai/moonshotai/Kimi-K2.7-Code", "together_ai/MiniMaxAI/MiniMax-M3", "together_ai/thinkingmachines/Inkling", "together_ai/thinkingmachines/Inkling-Small", @@ -27,10 +25,8 @@ SERVERLESS_CHAT_MODELS: Final = ( "together_ai/Qwen/Qwen3.7-Plus", "together_ai/Qwen/Qwen3.6-Plus", "together_ai/Qwen/Qwen3.5-9B", - "together_ai/nvidia/nemotron-3-ultra-550b-a55b", "together_ai/meta-models/Muse-Glimmer-30B", "together_ai/google/gemma-4-31B-it", - "together_ai/pearl-ai/gemma-4-31b-it", "together_ai/arize-ai/qwen-2-1.5b-instruct", "together_ai/Prism-ML/Ternary-Bonsai-27B", "together_ai/openai/gpt-oss-120b", @@ -39,6 +35,10 @@ SERVERLESS_CHAT_MODELS: Final = ( ) DEPRECATED_MODELS: Final = { + "together_ai/nvidia/nemotron-3-ultra-550b-a55b": "2026-08-27", + "together_ai/pearl-ai/gemma-4-31b-it": "2026-08-27", + "together_ai/deepseek-ai/DeepSeek-V4-Pro": "2026-08-27", + "together_ai/moonshotai/Kimi-K2.7-Code": "2026-08-27", "together_ai/google/gemma-3n-E4B-it": "2026-08-25", "together_ai/meta-llama/Llama-Guard-4-12B": "2026-08-25", "together_ai/Qwen/Qwen3-235B-A22B-Instruct-2507-tput": "2026-07-10", From a0689f04c46d7db2d1ba11ca0a91026dc2609c55 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 09:48:19 -0700 Subject: [PATCH 32/44] fix(model_prices): cap ministral-3-3b at Mistral API's 131072 and mirror Anthropic family flags on new DeepInfra Claude rows --- litellm/model_prices_and_context_window_backup.json | 9 ++++++--- model_prices_and_context_window.json | 9 ++++++--- 2 files changed, 12 insertions(+), 6 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 0d867955544..b5d40e2800c 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -31558,9 +31558,9 @@ "mistral/ministral-3-3b-2512": { "input_cost_per_token": 1e-07, "litellm_provider": "mistral", - "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1e-07, "source": "https://mistral.ai/pricing", @@ -53130,10 +53130,12 @@ "output_cost_per_token": 1.5e-05, "litellm_provider": "deepinfra", "mode": "chat", + "prompt_cache_min_tokens": 1024, "supports_function_calling": true, "supports_tool_choice": true, "supports_response_schema": true, "supports_reasoning": true, + "supports_adaptive_thinking": true, "supports_vision": true, "source": "https://deepinfra.com/pricing" }, @@ -53920,6 +53922,7 @@ "output_cost_per_token": 5e-06, "litellm_provider": "deepinfra", "mode": "chat", + "prompt_cache_min_tokens": 4096, "supports_function_calling": true, "supports_tool_choice": true, "supports_response_schema": true, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 0d867955544..b5d40e2800c 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -31558,9 +31558,9 @@ "mistral/ministral-3-3b-2512": { "input_cost_per_token": 1e-07, "litellm_provider": "mistral", - "max_input_tokens": 262144, - "max_output_tokens": 262144, - "max_tokens": 262144, + "max_input_tokens": 131072, + "max_output_tokens": 131072, + "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1e-07, "source": "https://mistral.ai/pricing", @@ -53130,10 +53130,12 @@ "output_cost_per_token": 1.5e-05, "litellm_provider": "deepinfra", "mode": "chat", + "prompt_cache_min_tokens": 1024, "supports_function_calling": true, "supports_tool_choice": true, "supports_response_schema": true, "supports_reasoning": true, + "supports_adaptive_thinking": true, "supports_vision": true, "source": "https://deepinfra.com/pricing" }, @@ -53920,6 +53922,7 @@ "output_cost_per_token": 5e-06, "litellm_provider": "deepinfra", "mode": "chat", + "prompt_cache_min_tokens": 4096, "supports_function_calling": true, "supports_tool_choice": true, "supports_response_schema": true, From 5b80fb0fc0703fd7fbe48871c9c1b1de244d7b5b Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 12:30:13 -0700 Subject: [PATCH 33/44] fix(transcription): synthesize srt/vtt output for adapters without native subtitle formats Extract the Soniox SRT/VTT cue grouping and rendering into a shared litellm_core_utils/audio_utils/subtitle_utils module, have Gemini transcription request word timestamps whenever response_format is srt or vtt, and let the http handler rewrite the response text into the synthesized subtitle document (dropping the internally requested words array) for any provider config that opts in via supports_subtitle_synthesis --- .../audio_utils/subtitle_utils.py | 189 ++++++++++++++++++ .../audio_transcription/transformation.py | 10 + litellm/llms/custom_httpx/llm_http_handler.py | 17 +- .../audio_transcription/transformation.py | 16 +- litellm/llms/soniox/common_utils.py | 154 ++------------ .../audio_utils/__init__.py | 0 .../audio_utils/test_subtitle_utils.py | 134 +++++++++++++ .../custom_httpx/test_llm_http_handler.py | 29 +++ ...mini_audio_transcription_transformation.py | 75 +++++++ 9 files changed, 478 insertions(+), 146 deletions(-) create mode 100644 litellm/litellm_core_utils/audio_utils/subtitle_utils.py create mode 100644 tests/test_litellm/litellm_core_utils/audio_utils/__init__.py create mode 100644 tests/test_litellm/litellm_core_utils/audio_utils/test_subtitle_utils.py diff --git a/litellm/litellm_core_utils/audio_utils/subtitle_utils.py b/litellm/litellm_core_utils/audio_utils/subtitle_utils.py new file mode 100644 index 00000000000..025baed90a4 --- /dev/null +++ b/litellm/litellm_core_utils/audio_utils/subtitle_utils.py @@ -0,0 +1,189 @@ +"""Provider-agnostic SRT/WebVTT subtitle synthesis from timestamped transcription tokens.""" + +from collections.abc import Sequence +from dataclasses import dataclass +from functools import reduce +from typing import Final + +from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError + +CUE_MAX_TOKENS: Final = 15 +CUE_MAX_DURATION_MS: Final = 5000 + +SRT_RESPONSE_FORMAT: Final = "srt" +VTT_RESPONSE_FORMAT: Final = "vtt" +SUBTITLE_RESPONSE_FORMATS: Final = frozenset((SRT_RESPONSE_FORMAT, VTT_RESPONSE_FORMAT)) + + +@dataclass(frozen=True, slots=True) +class SubtitleToken: + text: str + start_ms: int | None = None + end_ms: int | None = None + speaker: str | int | None = None + + +@dataclass(frozen=True, slots=True) +class SubtitleCue: + start_ms: int + end_ms: int + text: str + + +@dataclass(frozen=True, slots=True) +class _CueAccumulator: + cues: tuple[SubtitleCue, ...] = () + texts: tuple[str, ...] = () + start_ms: int | None = None + end_ms: int | None = None + speaker: str | int | None = None + + +def _completed_cue(accumulator: _CueAccumulator) -> tuple[SubtitleCue, ...]: + if not accumulator.texts or accumulator.start_ms is None: + return () + text: Final = "".join(accumulator.texts).strip() + if not text: + return () + end_ms: Final = accumulator.end_ms if accumulator.end_ms is not None else accumulator.start_ms + return (SubtitleCue(start_ms=accumulator.start_ms, end_ms=end_ms, text=text),) + + +def _cue_break_reached(accumulator: _CueAccumulator, token: SubtitleToken) -> bool: + if len(accumulator.texts) >= CUE_MAX_TOKENS: + return True + return ( + accumulator.start_ms is not None + and token.start_ms is not None + and token.start_ms - accumulator.start_ms >= CUE_MAX_DURATION_MS + ) + + +def _absorb_token(accumulator: _CueAccumulator, token: SubtitleToken) -> _CueAccumulator: + if token.start_ms is None and accumulator.start_ms is None: + return accumulator + if token.speaker is not None and token.speaker != accumulator.speaker: + return _CueAccumulator( + cues=accumulator.cues + _completed_cue(accumulator), + texts=(token.text,), + start_ms=token.start_ms, + end_ms=token.end_ms, + speaker=token.speaker, + ) + if _cue_break_reached(accumulator, token): + return _CueAccumulator( + cues=accumulator.cues + _completed_cue(accumulator), + texts=(token.text,), + start_ms=token.start_ms, + end_ms=token.end_ms, + speaker=accumulator.speaker, + ) + return _CueAccumulator( + cues=accumulator.cues, + texts=(*accumulator.texts, token.text), + start_ms=accumulator.start_ms if accumulator.start_ms is not None else token.start_ms, + end_ms=token.end_ms if token.end_ms is not None else accumulator.end_ms, + speaker=accumulator.speaker, + ) + + +def group_subtitle_tokens_into_cues(tokens: Sequence[SubtitleToken]) -> tuple[SubtitleCue, ...]: + accumulator: Final = reduce(_absorb_token, tokens, _CueAccumulator()) + return accumulator.cues + _completed_cue(accumulator) + + +def _format_timestamp(total_ms: int, millis_separator: str) -> str: + clamped: Final = max(total_ms, 0) + hours, hour_remainder = divmod(clamped, 3_600_000) + minutes, minute_remainder = divmod(hour_remainder, 60_000) + seconds, millis = divmod(minute_remainder, 1_000) + return f"{hours:02d}:{minutes:02d}:{seconds:02d}{millis_separator}{millis:03d}" + + +def _render_srt(cues: Sequence[SubtitleCue]) -> str: + lines: Final = tuple( + line + for index, cue in enumerate(cues, start=1) + for line in ( + str(index), + f"{_format_timestamp(cue.start_ms, ',')} --> {_format_timestamp(cue.end_ms, ',')}", + cue.text, + "", + ) + ) + return "\n".join(lines) + + +def _render_vtt(cues: Sequence[SubtitleCue]) -> str: + cue_lines: Final = tuple( + line + for cue in cues + for line in ( + f"{_format_timestamp(cue.start_ms, '.')} --> {_format_timestamp(cue.end_ms, '.')}", + cue.text, + "", + ) + ) + return "\n".join(("WEBVTT", "", *cue_lines)) + + +def render_subtitle_tokens_as_srt(tokens: Sequence[SubtitleToken]) -> str: + """Render tokens as an SRT document; empty string when no token has timestamp data.""" + cues: Final = group_subtitle_tokens_into_cues(tokens) + if not cues: + return "" + return _render_srt(cues) + + +def render_subtitle_tokens_as_vtt(tokens: Sequence[SubtitleToken]) -> str: + """Render tokens as a WebVTT document; the WEBVTT header is emitted even without cues.""" + return _render_vtt(group_subtitle_tokens_into_cues(tokens)) + + +class TranscriptionWordTiming(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + word: str = "" + start: float | None = None + end: float | None = None + speaker: str | None = None + + +_WORD_TIMINGS_ADAPTER: Final = TypeAdapter(tuple[TranscriptionWordTiming, ...]) + + +def _seconds_to_ms(seconds: float | None) -> int | None: + if seconds is None: + return None + return round(seconds * 1000) + + +def _word_to_subtitle_token(word: TranscriptionWordTiming) -> SubtitleToken: + return SubtitleToken( + text=f"{word.word} ", + start_ms=_seconds_to_ms(word.start), + end_ms=_seconds_to_ms(word.end), + speaker=word.speaker, + ) + + +def _parse_word_timings(words: object) -> tuple[TranscriptionWordTiming, ...]: + try: + return _WORD_TIMINGS_ADAPTER.validate_python(words) + except ValidationError: + return () + + +def synthesize_subtitle_document(words: object, response_format: str) -> str | None: + """ + Build an SRT/VTT document from OpenAI verbose_json-style word dicts + (word/start/end in float seconds, optional speaker). Returns None when the + format is not a subtitle format or the words carry no usable timestamps. + """ + if response_format not in SUBTITLE_RESPONSE_FORMATS: + return None + tokens: Final = tuple(_word_to_subtitle_token(word) for word in _parse_word_timings(words)) + cues: Final = group_subtitle_tokens_into_cues(tokens) + if not cues: + return None + return _render_srt(cues) if response_format == SRT_RESPONSE_FORMAT else _render_vtt(cues) diff --git a/litellm/llms/base_llm/audio_transcription/transformation.py b/litellm/llms/base_llm/audio_transcription/transformation.py index 6d087102816..da1776d8dc7 100644 --- a/litellm/llms/base_llm/audio_transcription/transformation.py +++ b/litellm/llms/base_llm/audio_transcription/transformation.py @@ -40,6 +40,16 @@ class BaseAudioTranscriptionConfig(BaseConfig, ABC): def get_supported_openai_params(self, model: str) -> list[OpenAIAudioTranscriptionOptionalParams]: pass + @property + def supports_subtitle_synthesis(self) -> bool: + """ + Opt-in for providers without a native srt/vtt response body: when True + and the user asked for response_format srt/vtt, the http handler + synthesizes the subtitle document from the word timestamps the + provider's TranscriptionResponse carries in `words`. + """ + return False + def get_complete_url( self, api_base: str | None, diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 1e8a3f00986..d90f7bd4514 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -25,6 +25,7 @@ from litellm.litellm_core_utils.agentic_loop_settings import ( validated_max_agentic_loops, ) from litellm.litellm_core_utils.asyncify import run_async_function +from litellm.litellm_core_utils.audio_utils.subtitle_utils import synthesize_subtitle_document from litellm.litellm_core_utils.llm_request_utils import serialize_multipart_form_fields from litellm.litellm_core_utils.realtime_errors import realtime_error_event, websocket_close_reason from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming @@ -1296,9 +1297,23 @@ class BaseLLMHTTPHandler: api_key: str | None, ) -> TranscriptionResponse: """Shared logic for transforming audio transcription responses.""" - return provider_config.transform_audio_transcription_response( + transformed: Final = provider_config.transform_audio_transcription_response( raw_response=response, ) + if not provider_config.supports_subtitle_synthesis: + return transformed + requested_format: Final = optional_params.get("response_format") + if not isinstance(requested_format, str): + return transformed + document: Final = synthesize_subtitle_document( + words=transformed.get("words"), + response_format=requested_format, + ) + if document is None: + return transformed + transformed.text = document + delattr(transformed, "words") + return transformed def audio_transcriptions( self, diff --git a/litellm/llms/gemini/audio_transcription/transformation.py b/litellm/llms/gemini/audio_transcription/transformation.py index 8b7733fa3c8..85335371b2b 100644 --- a/litellm/llms/gemini/audio_transcription/transformation.py +++ b/litellm/llms/gemini/audio_transcription/transformation.py @@ -4,6 +4,7 @@ from typing import Final from httpx import Headers, Response +from litellm.litellm_core_utils.audio_utils.subtitle_utils import SUBTITLE_RESPONSE_FORMATS from litellm.litellm_core_utils.audio_utils.utils import ( normalize_transcription_language_to_bcp47, process_audio_file, @@ -48,6 +49,10 @@ class GeminiAudioTranscriptionConfig(BaseAudioTranscriptionConfig): ) -> list[OpenAIAudioTranscriptionOptionalParams]: # mutable-ok: BaseAudioTranscriptionConfig signature return ["language", "response_format", "timestamp_granularities"] # mutable-ok: base contract returns a list + @property + def supports_subtitle_synthesis(self) -> bool: + return True + def map_openai_params( self, non_default_params: Mapping[str, object], @@ -215,16 +220,17 @@ def _language_config(language: object) -> GeminiTranscriptionConfig: return language_config -def _timestamp_config(timestamp_granularities: object) -> GeminiTranscriptionConfig: - if isinstance(timestamp_granularities, list) and "word" in timestamp_granularities: - return _WORD_TIMESTAMP_CONFIG - return _EMPTY_TRANSCRIPTION_CONFIG +def _timestamp_config(timestamp_granularities: object, response_format: object) -> GeminiTranscriptionConfig: + wants_word_timestamps: Final = ( + isinstance(timestamp_granularities, list) and "word" in timestamp_granularities + ) or response_format in SUBTITLE_RESPONSE_FORMATS + return _WORD_TIMESTAMP_CONFIG if wants_word_timestamps else _EMPTY_TRANSCRIPTION_CONFIG def _build_transcription_config(optional_params: Mapping[str, object]) -> GeminiTranscriptionConfig: transcription_config: Final[GeminiTranscriptionConfig] = { **_language_config(optional_params.get("language")), - **_timestamp_config(optional_params.get("timestamp_granularities")), + **_timestamp_config(optional_params.get("timestamp_granularities"), optional_params.get("response_format")), } return transcription_config diff --git a/litellm/llms/soniox/common_utils.py b/litellm/llms/soniox/common_utils.py index 90be94b8133..e2b18736e0e 100644 --- a/litellm/llms/soniox/common_utils.py +++ b/litellm/llms/soniox/common_utils.py @@ -4,6 +4,11 @@ Shared utilities for the Soniox provider (https://soniox.com). from typing import Any, Final +from litellm.litellm_core_utils.audio_utils.subtitle_utils import ( + SubtitleToken, + render_subtitle_tokens_as_srt, + render_subtitle_tokens_as_vtt, +) from litellm.llms.base_llm.chat.transformation import BaseLLMException # Soniox API base URL. @@ -109,121 +114,13 @@ def render_soniox_tokens(tokens: list[dict[str, Any]]) -> str: return "".join(text_parts) -# --------------------------------------------------------------------------- -# SRT / VTT subtitle rendering -# --------------------------------------------------------------------------- - -# Maximum number of tokens to group into a single subtitle cue. -_CUE_MAX_TOKENS: Final[int] = 15 - -# Maximum duration (in ms) for a single cue before forcing a break. -_CUE_MAX_DURATION_MS: Final[int] = 5000 - - -def _format_timestamp_srt(ms: int) -> str: - """Format milliseconds as SRT timestamp: HH:MM:SS,mmm""" - ms = max(ms, 0) - hours: Final = ms // 3_600_000 - ms %= 3_600_000 - minutes: Final = ms // 60_000 - ms %= 60_000 - seconds: Final = ms // 1_000 - millis: Final = ms % 1_000 - return f"{hours:02d}:{minutes:02d}:{seconds:02d},{millis:03d}" - - -def _format_timestamp_vtt(ms: int) -> str: - """Format milliseconds as VTT timestamp: HH:MM:SS.mmm""" - ms = max(ms, 0) - hours: Final = ms // 3_600_000 - ms %= 3_600_000 - minutes: Final = ms // 60_000 - ms %= 60_000 - seconds: Final = ms // 1_000 - millis: Final = ms % 1_000 - return f"{hours:02d}:{minutes:02d}:{seconds:02d}.{millis:03d}" - - -def _group_tokens_into_cues( - tokens: list[dict[str, Any]], -) -> list[dict[str, Any]]: - """ - Group Soniox tokens into subtitle cues. - - Each cue has: - - start_ms: int - - end_ms: int - - text: str - - Grouping heuristics: - - A new cue starts when token count exceeds _CUE_MAX_TOKENS. - - A new cue starts when duration exceeds _CUE_MAX_DURATION_MS. - - A new cue starts when the speaker changes (if diarization is on). - - Tokens without timestamps are appended to the current cue. - """ - cues: Final[list[dict[str, Any]]] = [] - current_tokens: list[str] = [] - current_start: int | None = None - current_end: int | None = None - current_speaker: Any | None = None - - def _flush() -> None: - if current_tokens and current_start is not None: - text: Final = "".join(current_tokens).strip() - if text: - cues.append( - { - "start_ms": current_start, - "end_ms": (current_end if current_end is not None else current_start), - "text": text, - } - ) - - for token in tokens: - start_ms = token.get("start_ms") - end_ms = token.get("end_ms") - text = token.get("text", "") - speaker = token.get("speaker") - - # Skip tokens with no timestamp data entirely if we have no cue started - if start_ms is None and current_start is None: - continue - - # Speaker change forces a new cue - if speaker is not None and speaker != current_speaker: - _flush() - current_tokens = [] - current_start = start_ms - current_end = end_ms - current_speaker = speaker - current_tokens.append(text) - continue - - # Duration or token count exceeded -> flush - should_break = False - if ( - len(current_tokens) >= _CUE_MAX_TOKENS - or current_start is not None - and start_ms is not None - and (start_ms - current_start) >= _CUE_MAX_DURATION_MS - ): - should_break = True - - if should_break: - _flush() - current_tokens = [] - current_start = start_ms - current_end = end_ms - current_tokens.append(text) - else: - if current_start is None: - current_start = start_ms - if end_ms is not None: - current_end = end_ms - current_tokens.append(text) - - _flush() - return cues +def _soniox_token_to_subtitle_token(token: dict[str, Any]) -> SubtitleToken: + return SubtitleToken( + text=token.get("text", ""), + start_ms=token.get("start_ms"), + end_ms=token.get("end_ms"), + speaker=token.get("speaker"), + ) def render_soniox_tokens_as_srt(tokens: list[dict[str, Any]]) -> str: @@ -232,20 +129,7 @@ def render_soniox_tokens_as_srt(tokens: list[dict[str, Any]]) -> str: Returns an empty string if no tokens have timestamp data. """ - cues: Final = _group_tokens_into_cues(tokens) - if not cues: - return "" - - lines: Final[list[str]] = [] - for idx, cue in enumerate(cues, start=1): - start = _format_timestamp_srt(cue["start_ms"]) - end = _format_timestamp_srt(cue["end_ms"]) - lines.append(str(idx)) - lines.append(f"{start} --> {end}") - lines.append(cue["text"]) - lines.append("") # blank line between cues - - return "\n".join(lines) + return render_subtitle_tokens_as_srt(tuple(_soniox_token_to_subtitle_token(token) for token in tokens)) def render_soniox_tokens_as_vtt(tokens: list[dict[str, Any]]) -> str: @@ -254,14 +138,4 @@ def render_soniox_tokens_as_vtt(tokens: list[dict[str, Any]]) -> str: Returns the VTT header even if no cues are present. """ - cues: Final = _group_tokens_into_cues(tokens) - - lines: Final[list[str]] = ["WEBVTT", ""] - for cue in cues: - start = _format_timestamp_vtt(cue["start_ms"]) - end = _format_timestamp_vtt(cue["end_ms"]) - lines.append(f"{start} --> {end}") - lines.append(cue["text"]) - lines.append("") # blank line between cues - - return "\n".join(lines) + return render_subtitle_tokens_as_vtt(tuple(_soniox_token_to_subtitle_token(token) for token in tokens)) diff --git a/tests/test_litellm/litellm_core_utils/audio_utils/__init__.py b/tests/test_litellm/litellm_core_utils/audio_utils/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/test_litellm/litellm_core_utils/audio_utils/test_subtitle_utils.py b/tests/test_litellm/litellm_core_utils/audio_utils/test_subtitle_utils.py new file mode 100644 index 00000000000..dcc4163ff10 --- /dev/null +++ b/tests/test_litellm/litellm_core_utils/audio_utils/test_subtitle_utils.py @@ -0,0 +1,134 @@ +from litellm.litellm_core_utils.audio_utils.subtitle_utils import ( + SubtitleToken, + render_subtitle_tokens_as_srt, + render_subtitle_tokens_as_vtt, + synthesize_subtitle_document, +) + + +class TestRenderSubtitleTokensAsSrt: + def test_single_cue_full_document(self): + tokens = ( + SubtitleToken(text="Hello ", start_ms=0, end_ms=500), + SubtitleToken(text="world.", start_ms=500, end_ms=1000), + ) + assert render_subtitle_tokens_as_srt(tokens) == "1\n00:00:00,000 --> 00:00:01,000\nHello world.\n" + + def test_speaker_change_starts_a_new_cue(self): + tokens = ( + SubtitleToken(text="Hi.", start_ms=0, end_ms=1000, speaker="spk:0"), + SubtitleToken(text="Hey.", start_ms=1500, end_ms=2500, speaker="spk:1"), + ) + assert render_subtitle_tokens_as_srt(tokens) == ( + "1\n00:00:00,000 --> 00:00:01,000\nHi.\n\n2\n00:00:01,500 --> 00:00:02,500\nHey.\n" + ) + + def test_token_cap_starts_a_new_cue_after_15_tokens(self): + tokens = tuple( + SubtitleToken(text=f"{index} ", start_ms=index * 100, end_ms=index * 100 + 100) for index in range(16) + ) + assert render_subtitle_tokens_as_srt(tokens) == ( + "1\n00:00:00,000 --> 00:00:01,500\n0 1 2 3 4 5 6 7 8 9 10 11 12 13 14\n" + "\n2\n00:00:01,500 --> 00:00:01,600\n15\n" + ) + + def test_duration_cap_starts_a_new_cue_at_5000ms(self): + tokens = ( + SubtitleToken(text="Alpha ", start_ms=0, end_ms=400), + SubtitleToken(text="beta ", start_ms=2000, end_ms=2400), + SubtitleToken(text="gamma.", start_ms=5000, end_ms=5400), + ) + assert render_subtitle_tokens_as_srt(tokens) == ( + "1\n00:00:00,000 --> 00:00:02,400\nAlpha beta\n\n2\n00:00:05,000 --> 00:00:05,400\ngamma.\n" + ) + + def test_timestampless_token_joins_the_current_cue(self): + tokens = ( + SubtitleToken(text="Hello ", start_ms=0, end_ms=500), + SubtitleToken(text="there "), + SubtitleToken(text="world.", start_ms=900, end_ms=1300), + ) + assert render_subtitle_tokens_as_srt(tokens) == "1\n00:00:00,000 --> 00:00:01,300\nHello there world.\n" + + def test_only_timestampless_tokens_renders_empty(self): + assert render_subtitle_tokens_as_srt((SubtitleToken(text="no timestamps"),)) == "" + + def test_empty_tokens_render_empty(self): + assert render_subtitle_tokens_as_srt(()) == "" + + def test_timestamps_past_one_hour(self): + tokens = (SubtitleToken(text="Late.", start_ms=3_661_001, end_ms=3_662_002),) + assert render_subtitle_tokens_as_srt(tokens) == "1\n01:01:01,001 --> 01:01:02,002\nLate.\n" + + def test_negative_timestamps_clamp_to_zero(self): + tokens = (SubtitleToken(text="Early.", start_ms=-100, end_ms=-50),) + assert render_subtitle_tokens_as_srt(tokens) == "1\n00:00:00,000 --> 00:00:00,000\nEarly.\n" + + def test_missing_end_falls_back_to_cue_start(self): + tokens = (SubtitleToken(text="Open.", start_ms=1200),) + assert render_subtitle_tokens_as_srt(tokens) == "1\n00:00:01,200 --> 00:00:01,200\nOpen.\n" + + +class TestRenderSubtitleTokensAsVtt: + def test_single_cue_full_document(self): + tokens = ( + SubtitleToken(text="Hello ", start_ms=0, end_ms=500), + SubtitleToken(text="world.", start_ms=500, end_ms=1000), + ) + assert render_subtitle_tokens_as_vtt(tokens) == "WEBVTT\n\n00:00:00.000 --> 00:00:01.000\nHello world.\n" + + def test_empty_tokens_render_header_only(self): + assert render_subtitle_tokens_as_vtt(()) == "WEBVTT\n" + + def test_timestamps_past_one_hour_use_dot_separator(self): + tokens = (SubtitleToken(text="Late.", start_ms=3_661_001, end_ms=3_662_002),) + assert render_subtitle_tokens_as_vtt(tokens) == "WEBVTT\n\n01:01:01.001 --> 01:01:02.002\nLate.\n" + + def test_speaker_change_starts_a_new_cue(self): + tokens = ( + SubtitleToken(text="Hi.", start_ms=0, end_ms=1000, speaker=1), + SubtitleToken(text="Hey.", start_ms=1500, end_ms=2500, speaker=2), + ) + assert render_subtitle_tokens_as_vtt(tokens) == ( + "WEBVTT\n\n00:00:00.000 --> 00:00:01.000\nHi.\n\n00:00:01.500 --> 00:00:02.500\nHey.\n" + ) + + +class TestSynthesizeSubtitleDocument: + WORDS = [ + {"word": "Four", "start": 0.4, "end": 0.7, "speaker": "spk:0"}, + {"word": "score", "start": 0.7, "end": 1.1, "speaker": "spk:0"}, + ] + + def test_srt_from_words_converts_seconds_to_milliseconds(self): + assert synthesize_subtitle_document(self.WORDS, "srt") == "1\n00:00:00,400 --> 00:00:01,100\nFour score\n" + + def test_vtt_from_words_converts_seconds_to_milliseconds(self): + assert synthesize_subtitle_document(self.WORDS, "vtt") == ( + "WEBVTT\n\n00:00:00.400 --> 00:00:01.100\nFour score\n" + ) + + def test_speaker_change_splits_cues(self): + words = [ + {"word": "Hi", "start": 0.0, "end": 0.5, "speaker": "spk:0"}, + {"word": "Hey", "start": 0.6, "end": 1.0, "speaker": "spk:1"}, + ] + assert synthesize_subtitle_document(words, "srt") == ( + "1\n00:00:00,000 --> 00:00:00,500\nHi\n\n2\n00:00:00,600 --> 00:00:01,000\nHey\n" + ) + + def test_non_subtitle_format_returns_none(self): + assert synthesize_subtitle_document(self.WORDS, "verbose_json") is None + assert synthesize_subtitle_document(self.WORDS, "json") is None + + def test_missing_words_returns_none(self): + assert synthesize_subtitle_document(None, "srt") is None + assert synthesize_subtitle_document([], "srt") is None + + def test_words_without_timestamps_return_none(self): + assert synthesize_subtitle_document([{"word": "Hello"}], "srt") is None + assert synthesize_subtitle_document([{"word": "Hello"}], "vtt") is None + + def test_malformed_words_return_none(self): + assert synthesize_subtitle_document("not words", "srt") is None + assert synthesize_subtitle_document([{"word": "ok", "start": "not-a-number"}], "srt") is None diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py index 18d1aa949a8..4eaa18b5aa9 100644 --- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py @@ -1901,6 +1901,35 @@ async def test_async_audio_transcriptions_sends_dict_data_as_json_body(): assert response.text == "transcribed" +class _WordTimestampAudioTranscriptionConfig(_JSONBodyAudioTranscriptionConfig): + def transform_audio_transcription_response(self, raw_response): + payload = raw_response.json() + response = TranscriptionResponse(text=payload["text"]) + response["words"] = payload["words"] + return response + + +def test_transform_audio_transcription_response_without_subtitle_opt_in_keeps_text_and_words(): + words = [ + {"word": "hello", "start": 0.0, "end": 0.5}, + {"word": "world", "start": 0.5, "end": 1.0}, + ] + raw_response = httpx.Response(200, json={"text": "hello world", "words": words}) + + response = BaseLLMHTTPHandler()._transform_audio_transcription_response( + provider_config=_WordTimestampAudioTranscriptionConfig(), + model="test-model", + response=raw_response, + model_response=TranscriptionResponse(), + logging_obj=Mock(), + optional_params={"response_format": "srt"}, + api_key=None, + ) + + assert response.text == "hello world" + assert response["words"] == words + + @pytest.mark.asyncio async def test_async_retrieve_file_content_raises_on_http_error(): """ diff --git a/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py b/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py index fef037974a7..0a06fb968c4 100644 --- a/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py +++ b/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py @@ -169,6 +169,33 @@ class TestTransformRequest: } } + @pytest.mark.parametrize("response_format", ["srt", "vtt"]) + def test_subtitle_response_format_requests_word_timestamps(self, config, response_format): + request_data = config.transform_audio_transcription_request( + model="gemini-3.5-transcribe", + audio_file=("sample.wav", AUDIO_BYTES, "audio/wav"), + optional_params={"response_format": response_format}, + litellm_params={}, + ) + transcription_config = request_data.data["generation_config"]["transcription_config"] + assert json.loads(json.dumps(transcription_config)) == { + "mode": { + "type": "verbatim", + "timestamp_granularities": ["word"], + "diarization_mode": "speaker", + } + } + + @pytest.mark.parametrize("response_format", ["json", "text", "verbose_json"]) + def test_non_subtitle_response_format_sends_no_mode(self, config, response_format): + request_data = config.transform_audio_transcription_request( + model="gemini-3.5-transcribe", + audio_file=("sample.wav", AUDIO_BYTES, "audio/wav"), + optional_params={"response_format": response_format}, + litellm_params={}, + ) + assert "generation_config" not in request_data.data + def test_segment_granularity_sends_no_mode(self, config): request_data = config.transform_audio_transcription_request( model="gemini-3.5-transcribe", @@ -214,6 +241,54 @@ class TestTransformResponse: assert response.get("duration") is None +class TestSubtitleSynthesisThroughHandler: + def _transform(self, config, response_format): + from unittest.mock import Mock + + from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler + from litellm.types.utils import TranscriptionResponse + + return BaseLLMHTTPHandler()._transform_audio_transcription_response( + provider_config=config, + model="gemini-3.5-transcribe", + response=make_response(COMPLETED_RESPONSE), + model_response=TranscriptionResponse(), + logging_obj=Mock(), + optional_params={"response_format": response_format}, + api_key=None, + ) + + def test_supports_subtitle_synthesis(self, config): + assert config.supports_subtitle_synthesis is True + + def test_srt_synthesizes_subtitle_document_and_drops_words(self, config): + response = self._transform(config, "srt") + assert response.text == ( + "1\n00:00:00,100 --> 00:00:00,400\nHello\n\n2\n00:00:00,500 --> 00:00:00,900\nworld.\n" + ) + assert "words" not in response + assert response["task"] == "transcribe" + assert response["duration"] == 0.9 + assert response.usage.total_tokens == 200 + + def test_vtt_synthesizes_subtitle_document_and_drops_words(self, config): + response = self._transform(config, "vtt") + assert response.text == ( + "WEBVTT\n\n00:00:00.100 --> 00:00:00.400\nHello\n\n00:00:00.500 --> 00:00:00.900\nworld.\n" + ) + assert "words" not in response + assert response.usage.total_tokens == 200 + + @pytest.mark.parametrize("response_format", ["json", "verbose_json"]) + def test_non_subtitle_formats_keep_plain_text_and_words(self, config, response_format): + response = self._transform(config, response_format) + assert response.text == "Hello world." + assert response["words"] == [ + {"word": "Hello", "start": 0.1, "end": 0.4, "speaker": "spk:0"}, + {"word": "world.", "start": 0.5, "end": 0.9, "speaker": "spk:1"}, + ] + + class TestCostRegression: @pytest.fixture def local_cost_map(self, monkeypatch): From 5ea81a0ae0391b2f2e337ea92343221aa54008b0 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 12:58:20 -0700 Subject: [PATCH 34/44] perf(subtitle_utils): make cue grouping linear in cue count --- .../audio_utils/subtitle_utils.py | 28 +++++++++++-------- 1 file changed, 16 insertions(+), 12 deletions(-) diff --git a/litellm/litellm_core_utils/audio_utils/subtitle_utils.py b/litellm/litellm_core_utils/audio_utils/subtitle_utils.py index 025baed90a4..615873e295d 100644 --- a/litellm/litellm_core_utils/audio_utils/subtitle_utils.py +++ b/litellm/litellm_core_utils/audio_utils/subtitle_utils.py @@ -2,7 +2,7 @@ from collections.abc import Sequence from dataclasses import dataclass -from functools import reduce +from itertools import accumulate, chain from typing import Final from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError @@ -32,7 +32,6 @@ class SubtitleCue: @dataclass(frozen=True, slots=True) class _CueAccumulator: - cues: tuple[SubtitleCue, ...] = () texts: tuple[str, ...] = () start_ms: int | None = None end_ms: int | None = None @@ -59,27 +58,27 @@ def _cue_break_reached(accumulator: _CueAccumulator, token: SubtitleToken) -> bo ) -def _absorb_token(accumulator: _CueAccumulator, token: SubtitleToken) -> _CueAccumulator: +_AbsorbStep = tuple[tuple[SubtitleCue, ...], _CueAccumulator] + + +def _absorb_token(accumulator: _CueAccumulator, token: SubtitleToken) -> _AbsorbStep: if token.start_ms is None and accumulator.start_ms is None: - return accumulator + return (), accumulator if token.speaker is not None and token.speaker != accumulator.speaker: - return _CueAccumulator( - cues=accumulator.cues + _completed_cue(accumulator), + return _completed_cue(accumulator), _CueAccumulator( texts=(token.text,), start_ms=token.start_ms, end_ms=token.end_ms, speaker=token.speaker, ) if _cue_break_reached(accumulator, token): - return _CueAccumulator( - cues=accumulator.cues + _completed_cue(accumulator), + return _completed_cue(accumulator), _CueAccumulator( texts=(token.text,), start_ms=token.start_ms, end_ms=token.end_ms, speaker=accumulator.speaker, ) - return _CueAccumulator( - cues=accumulator.cues, + return (), _CueAccumulator( texts=(*accumulator.texts, token.text), start_ms=accumulator.start_ms if accumulator.start_ms is not None else token.start_ms, end_ms=token.end_ms if token.end_ms is not None else accumulator.end_ms, @@ -87,9 +86,14 @@ def _absorb_token(accumulator: _CueAccumulator, token: SubtitleToken) -> _CueAcc ) +def _absorb_step(carry: _AbsorbStep, token: SubtitleToken) -> _AbsorbStep: + return _absorb_token(carry[1], token) + + def group_subtitle_tokens_into_cues(tokens: Sequence[SubtitleToken]) -> tuple[SubtitleCue, ...]: - accumulator: Final = reduce(_absorb_token, tokens, _CueAccumulator()) - return accumulator.cues + _completed_cue(accumulator) + steps: Final = tuple(accumulate(tokens, _absorb_step, initial=((), _CueAccumulator()))) + completed: Final = chain.from_iterable(emitted for emitted, _ in steps) + return (*completed, *_completed_cue(steps[-1][1])) def _format_timestamp(total_ms: int, millis_separator: str) -> str: From f2e64d9818322ab617b98f13099e9b0cf417423d Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 13:15:17 -0700 Subject: [PATCH 35/44] fix(gemini): ignore non-string response_format when deciding on word timestamps --- .../llms/gemini/audio_transcription/transformation.py | 2 +- .../test_gemini_audio_transcription_transformation.py | 9 +++++++++ .../test_soniox_audio_transcription_transformation.py | 4 ++-- 3 files changed, 12 insertions(+), 3 deletions(-) diff --git a/litellm/llms/gemini/audio_transcription/transformation.py b/litellm/llms/gemini/audio_transcription/transformation.py index 85335371b2b..c8dd7a9a5ff 100644 --- a/litellm/llms/gemini/audio_transcription/transformation.py +++ b/litellm/llms/gemini/audio_transcription/transformation.py @@ -223,7 +223,7 @@ def _language_config(language: object) -> GeminiTranscriptionConfig: def _timestamp_config(timestamp_granularities: object, response_format: object) -> GeminiTranscriptionConfig: wants_word_timestamps: Final = ( isinstance(timestamp_granularities, list) and "word" in timestamp_granularities - ) or response_format in SUBTITLE_RESPONSE_FORMATS + ) or (isinstance(response_format, str) and response_format in SUBTITLE_RESPONSE_FORMATS) return _WORD_TIMESTAMP_CONFIG if wants_word_timestamps else _EMPTY_TRANSCRIPTION_CONFIG diff --git a/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py b/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py index 0a06fb968c4..8b48ac0b467 100644 --- a/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py +++ b/tests/test_litellm/llms/gemini/audio_transcription/test_gemini_audio_transcription_transformation.py @@ -196,6 +196,15 @@ class TestTransformRequest: ) assert "generation_config" not in request_data.data + def test_non_string_response_format_sends_no_mode(self, config): + request_data = config.transform_audio_transcription_request( + model="gemini-3.5-transcribe", + audio_file=("sample.wav", AUDIO_BYTES, "audio/wav"), + optional_params={"response_format": {"type": "json_object"}}, + litellm_params={}, + ) + assert "generation_config" not in request_data.data + def test_segment_granularity_sends_no_mode(self, config): request_data = config.transform_audio_transcription_request( model="gemini-3.5-transcribe", diff --git a/tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py b/tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py index 7ee816d5d9e..261efcb7b24 100644 --- a/tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py +++ b/tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py @@ -477,12 +477,12 @@ class TestBuildResponseWithResponseFormat: } } # SRT requested but tokens have no start_ms/end_ms -> empty SRT - # falls back gracefully since _group_tokens_into_cues skips them + # falls back gracefully since group_subtitle_tokens_into_cues skips them resp = cfg._build_response_from_payload(payload, response_format="srt") # With no timestamp data, SRT rendering produces empty string, # but we still get output because the code checks `tokens` truthiness # before choosing SRT path. Actually the tokens list is truthy but - # _group_tokens_into_cues will produce no cues -> empty SRT string. + # group_subtitle_tokens_into_cues will produce no cues -> empty SRT string. # Let's verify it doesn't crash. assert isinstance(resp.text, str) From 474fbea81fdc1e9c0e103584bedde59c9c6fd379 Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Thu, 27 Aug 2026 13:20:28 -0700 Subject: [PATCH 36/44] test(e2e): let the together tool tests accept parallel calls The together backend is picked as the cheapest chat row that supports both tools and reasoning, which currently resolves to together_ai/openai/gpt-oss-120b. That row is marked supports_parallel_function_calling, so one weather prompt can legitimately come back as several get_weather calls. Both tool tests asserted exactly one call, so a parallel answer failed them even though the gateway handled it correctly. They now check every returned call instead of counting them: each one has to be a get_weather naming Paris, with an id a tool result can answer. Dropping, misnaming, or mangling a call is still red; only the count is the model's business. The round trips answer every call rather than just the first, which is also what the Anthropic Messages spec asks for. --- .../llm_translation/test_together_ai_e2e.py | 54 ++++++++++++------- 1 file changed, 36 insertions(+), 18 deletions(-) diff --git a/tests/e2e/llm_translation/test_together_ai_e2e.py b/tests/e2e/llm_translation/test_together_ai_e2e.py index 788b1858b73..a3ca7f0fc3b 100644 --- a/tests/e2e/llm_translation/test_together_ai_e2e.py +++ b/tests/e2e/llm_translation/test_together_ai_e2e.py @@ -210,16 +210,24 @@ def _deltas(result: StreamingResponse) -> list[_StreamDelta]: ] -def _single_weather_call(message: OutMessage) -> ToolCall: - assert message.tool_calls, f"Together dropped the tool call: {message}" - assert len(message.tool_calls) == 1, f"expected one tool call, got {message.tool_calls}" - call = message.tool_calls[0] +def _validated_weather_call_id(call: ToolCall) -> str: assert call.id, f"tool call carries no id, so a tool result cannot answer it: {call}" assert call.function.name == "get_weather", f"wrong tool called: {call}" assert call.function.arguments, f"tool call carries no arguments: {call}" args = _WeatherArgs.model_validate_json(call.function.arguments) assert "paris" in args.location.lower(), f"tool arguments lost the location: {args}" - return call + return call.id + + +def _weather_call_ids(message: OutMessage) -> tuple[str, ...]: + """The id of every tool call the model made, each one checked for the fields a + caller needs to answer it. The backend is whichever together_ai row is cheapest + with tools and reasoning, and those rows carry supports_parallel_function_calling, + so one weather prompt can legitimately come back as several get_weather calls. + What the gateway owes us is that each call survives translation intact; how many + the model chose to make is the model's business.""" + assert message.tool_calls, f"Together dropped the tool call: {message}" + return tuple(_validated_weather_call_id(call) for call in message.tool_calls) def _weather_call(client: PassthroughClient, key: str, model: str) -> OutMessage: @@ -289,7 +297,7 @@ class TestTogetherChatCompletions: self, client: PassthroughClient, resources: ResourceManager, reasoning_tool_backend: str ) -> None: model, key = _register(client, resources, reasoning_tool_backend) - _single_weather_call(_weather_call(client, key, model)) + _ = _weather_call_ids(_weather_call(client, key, model)) @pytest.mark.covers("llm.chat_completions.together_ai.tool_use.stream.works") def test_tool_call_is_streamed( @@ -328,8 +336,7 @@ class TestTogetherChatCompletions: ) -> None: model, key = _register(client, resources, reasoning_tool_backend) first = _weather_call(client, key, model) - call = _single_weather_call(first) - assert call.id is not None + call_ids = _weather_call_ids(first) answer = _message( unwrap( @@ -344,7 +351,10 @@ class TestTogetherChatCompletions: reasoning_content=first.reasoning_content, tool_calls=first.tool_calls, ), - ChatToolResultTurn(tool_call_id=call.id, content=WEATHER_REPORT), + *( + ChatToolResultTurn(tool_call_id=call_id, content=WEATHER_REPORT) + for call_id in call_ids + ), ], tools=[WEATHER_TOOL], max_tokens=512, @@ -470,9 +480,18 @@ def _tool_use_blocks(content: list[AnthropicContentBlock] | None) -> list[Anthro return [block for block in content if block.type == "tool_use"] +def _validated_tool_use_id(block: AnthropicContentBlock) -> str: + assert block.name == "get_weather", f"wrong tool called: {block}" + assert block.id, f"tool_use block carries no id, so a tool_result cannot answer it: {block}" + return block.id + + def _messages_weather_call( client: PassthroughClient, key: str, model: str -) -> tuple[list[AnthropicContentBlock], AnthropicContentBlock]: +) -> tuple[list[AnthropicContentBlock], tuple[str, ...]]: + """The blocks /v1/messages returned and the id of every tool_use among them. The + count is the model's choice (see _weather_call_ids); what this surface owes us is + that each tool_use arrives named and addressable.""" response = unwrap( client.proxy.messages( key, @@ -485,12 +504,9 @@ def _messages_weather_call( ) ) tool_uses = _tool_use_blocks(response.content) - assert len(tool_uses) == 1, f"expected one tool_use block, got {response.content}" - block = tool_uses[0] - assert block.name == "get_weather", f"wrong tool called: {block}" - assert block.id, f"tool_use block carries no id, so a tool_result cannot answer it: {block}" + assert tool_uses, f"/v1/messages carried no tool_use block: {response.content}" assert response.content is not None - return response.content, block + return response.content, tuple(_validated_tool_use_id(block) for block in tool_uses) class TestTogetherMessages: @@ -506,8 +522,7 @@ class TestTogetherMessages: self, client: PassthroughClient, resources: ResourceManager, reasoning_tool_backend: str ) -> None: model, key = _register(client, resources, reasoning_tool_backend) - first_content, block = _messages_weather_call(client, key, model) - assert block.id is not None + first_content, tool_use_ids = _messages_weather_call(client, key, model) response = unwrap( client.proxy.messages( @@ -520,7 +535,10 @@ class TestTogetherMessages: ChatMessage(role="user", content=WEATHER_PROMPT), AnthropicAssistantTurn(content=first_content), AnthropicToolResultTurn( - content=[AnthropicToolResultBlock(tool_use_id=block.id, content=WEATHER_REPORT)] + content=[ + AnthropicToolResultBlock(tool_use_id=tool_use_id, content=WEATHER_REPORT) + for tool_use_id in tool_use_ids + ] ), ], ), From 2cab010c73a6befe3e403a6eeb3d33df1ac6f897 Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Thu, 27 Aug 2026 13:34:09 -0700 Subject: [PATCH 37/44] test(e2e): check the tool input on the messages path too The /v1/messages validator checked a tool_use block's name and id but not its input, so a block whose location came back empty or wrong still passed, while the chat side rejected the same damage. That gap predates this branch; it is worth closing here because the point of the change is that every parallel call is checked rather than counted. AnthropicContentBlock now declares input as a typed field. It already survived on extra="allow", but reaching it from a test needs a real field to keep the e2e basedpyright gate at zero. Serialization is unchanged: bodies are dumped with exclude_none, so a block without an input still replays exactly as before. --- tests/e2e/llm_translation/test_together_ai_e2e.py | 3 +++ tests/e2e/models.py | 1 + 2 files changed, 4 insertions(+) diff --git a/tests/e2e/llm_translation/test_together_ai_e2e.py b/tests/e2e/llm_translation/test_together_ai_e2e.py index a3ca7f0fc3b..90581aadd43 100644 --- a/tests/e2e/llm_translation/test_together_ai_e2e.py +++ b/tests/e2e/llm_translation/test_together_ai_e2e.py @@ -483,6 +483,9 @@ def _tool_use_blocks(content: list[AnthropicContentBlock] | None) -> list[Anthro def _validated_tool_use_id(block: AnthropicContentBlock) -> str: assert block.name == "get_weather", f"wrong tool called: {block}" assert block.id, f"tool_use block carries no id, so a tool_result cannot answer it: {block}" + assert block.input is not None, f"tool_use block carries no input: {block}" + args = _WeatherArgs.model_validate(block.input) + assert "paris" in args.location.lower(), f"tool input lost the location: {args}" return block.id diff --git a/tests/e2e/models.py b/tests/e2e/models.py index 8d1b17ca256..e8a392599fd 100644 --- a/tests/e2e/models.py +++ b/tests/e2e/models.py @@ -421,6 +421,7 @@ class AnthropicContentBlock(BaseModel): text: str | None = None id: str | None = None name: str | None = None + input: dict[str, object] | None = None class AnthropicToolResultBlock(BaseModel): From 5d8eb049b5b1a514167207a001490e7b3398b863 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 13:36:31 -0700 Subject: [PATCH 38/44] fix(transcription): drop words from srt/vtt replies when synthesis falls back --- litellm/llms/custom_httpx/llm_http_handler.py | 15 ++++--- .../custom_httpx/test_llm_http_handler.py | 41 +++++++++++++++++++ 2 files changed, 50 insertions(+), 6 deletions(-) diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index d90f7bd4514..df5365017ac 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -25,7 +25,10 @@ from litellm.litellm_core_utils.agentic_loop_settings import ( validated_max_agentic_loops, ) from litellm.litellm_core_utils.asyncify import run_async_function -from litellm.litellm_core_utils.audio_utils.subtitle_utils import synthesize_subtitle_document +from litellm.litellm_core_utils.audio_utils.subtitle_utils import ( + SUBTITLE_RESPONSE_FORMATS, + synthesize_subtitle_document, +) from litellm.litellm_core_utils.llm_request_utils import serialize_multipart_form_fields from litellm.litellm_core_utils.realtime_errors import realtime_error_event, websocket_close_reason from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming @@ -1303,16 +1306,16 @@ class BaseLLMHTTPHandler: if not provider_config.supports_subtitle_synthesis: return transformed requested_format: Final = optional_params.get("response_format") - if not isinstance(requested_format, str): + if not isinstance(requested_format, str) or requested_format not in SUBTITLE_RESPONSE_FORMATS: return transformed document: Final = synthesize_subtitle_document( words=transformed.get("words"), response_format=requested_format, ) - if document is None: - return transformed - transformed.text = document - delattr(transformed, "words") + if document is not None: + transformed.text = document + if "words" in transformed: + delattr(transformed, "words") return transformed def audio_transcriptions( diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py index 4eaa18b5aa9..b37c0f466d2 100644 --- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py @@ -1930,6 +1930,47 @@ def test_transform_audio_transcription_response_without_subtitle_opt_in_keeps_te assert response["words"] == words +class _SubtitleSynthesisAudioTranscriptionConfig(_JSONBodyAudioTranscriptionConfig): + @property + def supports_subtitle_synthesis(self) -> bool: + return True + + def transform_audio_transcription_response(self, raw_response): + payload = raw_response.json() + response = TranscriptionResponse(text=payload["text"]) + if "words" in payload: + response["words"] = payload["words"] + return response + + +def _transform_subtitle_response(payload): + return BaseLLMHTTPHandler()._transform_audio_transcription_response( + provider_config=_SubtitleSynthesisAudioTranscriptionConfig(), + model="test-model", + response=httpx.Response(200, json=payload), + model_response=TranscriptionResponse(), + logging_obj=Mock(), + optional_params={"response_format": "srt"}, + api_key=None, + ) + + +def test_subtitle_synthesis_fallback_without_timings_drops_words(): + response = _transform_subtitle_response( + {"text": "hello world", "words": [{"word": "hello"}, {"word": "world"}]} + ) + + assert response.text == "hello world" + assert "words" not in response + + +def test_subtitle_synthesis_without_words_keeps_plain_text(): + response = _transform_subtitle_response({"text": "hello world"}) + + assert response.text == "hello world" + assert "words" not in response + + @pytest.mark.asyncio async def test_async_retrieve_file_content_raises_on_http_error(): """ From 88cb83484ba36d35f40f549525701f46c1769e34 Mon Sep 17 00:00:00 2001 From: yucheng-berri Date: Thu, 27 Aug 2026 13:41:30 -0700 Subject: [PATCH 39/44] fix(otel): anchor MCP tool-call spans to the gateway's own trace, link the client's context (#38317) Under otel_v2, a client that propagates W3C trace context in params._meta (SEP-414) pulled the tools/call span out of the gateway's trace: resolve_mcp_span_context parented the MCP span to the client's remote context and demoted the gateway's own transport span to a span link. The gateway's tracing backend only ever receives the gateway's half of such a trace, so the span was unreachable from the trace view and the POST transaction showed a dangling link. Invert the anchoring: the MCP tool-call and tools/list spans now always nest under the transport span of the request carrying the message, and the client's propagated context is recorded as the span link instead, so the correlation survives while every trace stays renderable. With no transport at all the span roots its own trace and still carries the link, keeping a single shape for the event. Both returned contexts are built on an explicitly empty base so ambient session state can never leak in, and the span inherits the transport's sampling decision like every other request-level span. --- litellm/integrations/otel/emitter.py | 6 +- litellm/integrations/otel/logger.py | 14 +-- litellm/integrations/otel/model/spans.py | 44 ++++----- litellm/integrations/otel/plumbing/context.py | 63 ++++++------ .../proxy/_experimental/mcp_server/server.py | 11 ++- .../integrations/otel/test_otel_v2_logger.py | 96 +++++++++++++++---- .../otel/test_otel_v2_sources_of_truth.py | 25 +++-- 7 files changed, 158 insertions(+), 101 deletions(-) diff --git a/litellm/integrations/otel/emitter.py b/litellm/integrations/otel/emitter.py index 244e58eddf3..101dbc6538d 100644 --- a/litellm/integrations/otel/emitter.py +++ b/litellm/integrations/otel/emitter.py @@ -146,7 +146,7 @@ class SpanEmitter: For callers that own and manage their own span lifecycle. ``tracer`` overrides the bound tracer for this span only, used for per-request multi-tenant credential routing. ``links`` records related-but-not-parent - spans (e.g. the transport span of an MCP message, per MCP semconv). + spans (e.g. the trace context an MCP client propagated in ``params._meta``). """ return (tracer or self._tracer).start_span( name, @@ -196,8 +196,8 @@ class SpanEmitter: Return the span, or ``None`` if it was deduplicated away. ``tracer`` overrides the bound tracer for this span, used for per-request routing. - ``links`` records related-but-not-parent spans (the transport span of an - MCP message). + ``links`` records related-but-not-parent spans (e.g. the trace context an + MCP client propagated in ``params._meta``). """ # LLM-call and MCP tool-call spans carry a dedup key (their request's # call id), so a sync+async double-firing coalesces. ``isinstance`` narrows diff --git a/litellm/integrations/otel/logger.py b/litellm/integrations/otel/logger.py index 4359b222d06..d2a32ef73b6 100644 --- a/litellm/integrations/otel/logger.py +++ b/litellm/integrations/otel/logger.py @@ -390,10 +390,10 @@ class OpenTelemetryV2(CustomLogger): MCP tool calls reach the success/failure callbacks like any other request (with ``call_type`` ``call_mcp_tool``), but they are not LLM calls and have - no ``pre_call`` carrier — so they get their own CLIENT span here. Per the MCP - semconv it parents to the trace context the client propagated in - ``params._meta`` (or starts a new root) and links the transport span, rather - than nesting under the HTTP/session span. Returns whether it handled the + no ``pre_call`` carrier — so they get their own CLIENT span here. It nests + under the transport span of the request carrying this message, and trace + context the client propagated in ``params._meta`` is recorded as a span + link (see ``resolve_mcp_span_context``). Returns whether it handled the event, so the caller skips the LLM-call path. The whole span is emitted at once (there is no boundary to open it at), deduped on the call id. """ @@ -436,9 +436,9 @@ class OpenTelemetryV2(CustomLogger): Like a tool call, listing reaches the success/failure callbacks (here with ``call_type`` ``list_mcp_tools``) with no ``pre_call`` carrier, so it gets its - own CLIENT span. Per the MCP semconv it parents to the ``params._meta`` trace - context (or starts a new root) and links the transport span, rather than - nesting under the HTTP/session span. Returns whether it handled the event so + own CLIENT span, nested under the transport span of the request carrying + this message with any ``params._meta`` trace context recorded as a span + link (see ``resolve_mcp_span_context``). Returns whether it handled the event so the caller skips the LLM-call path. """ raw_payload: Final = kwargs.get("standard_logging_object") diff --git a/litellm/integrations/otel/model/spans.py b/litellm/integrations/otel/model/spans.py index 08318f78b7c..35fc50a2a83 100644 --- a/litellm/integrations/otel/model/spans.py +++ b/litellm/integrations/otel/model/spans.py @@ -10,6 +10,8 @@ Canonical hierarchy:: │ └── DB_CALL (CLIENT) # its key/user/team lookups nest here ├── GUARDRAIL (INTERNAL) # request-lifecycle hook, sibling of LLM_CALL ├── LLM_CALL (CLIENT) + ├── MCP_TOOL_CALL (CLIENT) # nests under the POST carrying the message + ├── MCP_LIST_TOOLS (CLIENT) # (client-propagated context is a span link) └── DB_CALL (CLIENT) # e.g. the spend-log write Guardrails parent to PROXY_REQUEST, not LLM_CALL: pre/during/post-call guardrail @@ -18,14 +20,14 @@ before the LLM call even starts), so a guardrail is a sibling of the LLM call, not a child of it. The emitter parents every span to the ambient OTel context (the active server span), which matches this. -MCP spans (``MCP_TOOL_CALL``, ``MCP_LIST_TOOLS``) have two shapes, chosen at emit -time by :func:`resolve_mcp_span_context`. When the client propagates trace context -in ``params._meta`` MCP and the HTTP transport are independent contexts per the -OTel GenAI MCP semconv, so the span parents to that propagated context and records -the ``PROXY_REQUEST`` transport span as a span *link*, never a parent — the shape -this registry's ``parent=None, links=PROXY_REQUEST`` entry encodes. When nothing is -propagated (the common case) the span nests under the transport span of the request -carrying that message, so the tool call stays in one trace. +MCP spans (``MCP_TOOL_CALL``, ``MCP_LIST_TOOLS``) are parented at emit time by +:func:`resolve_mcp_span_context`: they nest under the ``PROXY_REQUEST`` transport +span of the request carrying that message, so the tool call stays in one trace. +Trace context the client propagated in ``params._meta`` (SEP-414) is recorded as +a span *link*, never the parent — a remote parent would root the span in a trace +whose root never reaches the gateway's tracing backend. Links always target that +remote client context, never a registry role, so ``SpanSpec`` declares no link +field; the concrete transport parent is resolved per message at emit time. Not every service call becomes a span — :func:`span_role_for_service` decides: @@ -85,25 +87,19 @@ class SpanSpec: role: SpanRole kind: LiteLLMSpanKind parent: SpanRole | None - links: SpanRole | None = None SPAN_REGISTRY: Final[dict[SpanRole, SpanSpec]] = { SpanRole.PROXY_REQUEST: SpanSpec(SpanRole.PROXY_REQUEST, LiteLLMSpanKind.SERVER, parent=None), SpanRole.LLM_CALL: SpanSpec(SpanRole.LLM_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST), # The proxy is an MCP client to the upstream server, so MCP spans are CLIENT - # spans. With trace context propagated in ``params._meta``, MCP and the HTTP - # transport are independent contexts (OTel GenAI MCP semconv): the span parents - # to the propagated context and records the PROXY_REQUEST transport span as a - # span *link*, never a parent — the shape ``parent=None, links=PROXY_REQUEST`` - # encodes. With nothing propagated, ``resolve_mcp_span_context`` nests the span - # under that message's transport span instead, keeping the call in one trace. - SpanRole.MCP_TOOL_CALL: SpanSpec( - SpanRole.MCP_TOOL_CALL, LiteLLMSpanKind.CLIENT, parent=None, links=SpanRole.PROXY_REQUEST - ), - SpanRole.MCP_LIST_TOOLS: SpanSpec( - SpanRole.MCP_LIST_TOOLS, LiteLLMSpanKind.CLIENT, parent=None, links=SpanRole.PROXY_REQUEST - ), + # spans. ``resolve_mcp_span_context`` nests them under the PROXY_REQUEST + # transport span of the request carrying that message (resolved per message at + # emit time), keeping the call in one trace. Trace context the client + # propagated in ``params._meta`` becomes a span *link* to that remote context, + # which is not a registry role, so ``SpanSpec`` has no link field. + SpanRole.MCP_TOOL_CALL: SpanSpec(SpanRole.MCP_TOOL_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST), + SpanRole.MCP_LIST_TOOLS: SpanSpec(SpanRole.MCP_LIST_TOOLS, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST), SpanRole.GUARDRAIL: SpanSpec(SpanRole.GUARDRAIL, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST), SpanRole.DB_CALL: SpanSpec(SpanRole.DB_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST), SpanRole.SERVICE: SpanSpec(SpanRole.SERVICE, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST), @@ -209,8 +205,8 @@ def service_span_name(data: "ServiceSpanData") -> str: def root_roles() -> list[SpanRole]: - """Roles with no in-process parent. They start a new trace unless they adopt a - remote parent (e.g. an MCP span joining the client's propagated context).""" + """Roles with no in-process parent, i.e. they start a new trace (only the + instrumentor-owned ``PROXY_REQUEST`` server span today).""" return [role for role, spec in SPAN_REGISTRY.items() if spec.parent is None] @@ -227,8 +223,6 @@ def validate_registry( raise ValueError(f"SPAN_REGISTRY[{role}] has mismatched role {spec.role}") if spec.parent is not None and spec.parent not in reg: raise ValueError(f"span role {role} declares unknown parent {spec.parent}") - if spec.links is not None and spec.links not in reg: - raise ValueError(f"span role {role} declares unknown link target {spec.links}") missing: Final = [role for role in SpanRole if role not in reg] if missing: raise ValueError(f"SPAN_REGISTRY is missing roles: {missing}") diff --git a/litellm/integrations/otel/plumbing/context.py b/litellm/integrations/otel/plumbing/context.py index 19b36c0b967..159a84b121f 100644 --- a/litellm/integrations/otel/plumbing/context.py +++ b/litellm/integrations/otel/plumbing/context.py @@ -57,8 +57,8 @@ def request_root_span() -> "Span | None": # The W3C trace-context carrier (``traceparent``/``tracestate``/``baggage``) the # MCP client propagated in the current request's ``params._meta``. The MCP gateway -# sets it per message so the MCP span can parent to the client's span rather than -# to the transport. A ``ContextVar`` because, like the root-span anchor, it must +# sets it per message so the MCP span can record the client's span as a span +# link. A ``ContextVar`` because, like the root-span anchor, it must # ride the request task and be readable by the inline success-logging callback. _mcp_message_trace_carrier: Final["ContextVar[Mapping[str, str] | None]"] = ContextVar( "litellm_otel_mcp_message_trace_carrier", default=None @@ -148,10 +148,10 @@ def _mcp_transport_span_context() -> "SpanContext | None": Prefers the transport the gateway published for this specific message; falls back to the ambient request anchor for paths that emit an MCP span on the - request task itself (the REST MCP endpoints, the SDK). Parenting and linking - only need the immutable context, and unlike ``mcp_message_transport_span`` they - stay correct against a transport that has already finished, so this does not - require the span to still be recording. + request task itself (the REST MCP endpoints). Parenting needs only the + immutable context, and unlike ``mcp_message_transport_span`` it stays correct + against a transport that has already finished, so this does not require the + span to still be recording. """ published: Final = _mcp_message_transport_span.get() if published is not None: @@ -222,25 +222,31 @@ def resolve_mcp_span_context( ) -> "tuple[Context, tuple[Link, ...]]": """Parent context + links for an MCP message span. + The span always nests under the transport span of the request carrying this + message, so a tool call and the ``POST`` that carried it stay in one trace. + The transport comes from :func:`_mcp_transport_span_context`, which is the + *current message's* POST rather than whatever request happened to open the + session, so a long-lived session does not glue every message under its first + request. + When the client propagates W3C trace context in the request's ``params._meta`` - (SEP-414), MCP and the underlying transport are independent lifecycles — one - streamable-HTTP session multiplexes many messages, and the client's own span is - the truthful parent. So, per the OTel GenAI MCP semconv: + (SEP-414), that remote context is recorded as a span *link*, never the parent. + The OTel GenAI MCP semconv prefers the inverse (remote parent, transport link), + but the gateway's tracing backend only ever receives the gateway's half of such + a trace: parenting into the client's trace id roots the span in a trace whose + root span never reaches the backend, so the span is unreachable from the trace + view and the transport transaction shows a dangling link (observed with + clients that propagate synthetic trace ids). Anchoring to the gateway's own + request and linking the client's context keeps every trace renderable while + preserving the client-side correlation. - * parent to the trace context the client propagated (a *remote* parent), and - * record the transport span as a *link*, never the parent. - - Almost no client implements SEP-414 yet, so in practice nothing is propagated. - Rooting the span there splits a single tool call into two disconnected traces - joined only by a link, which is how it surfaces in APM: the ``POST`` transaction - and the ``tools/call`` span share no trace. With no remote parent to honor, - parent to the transport span of the request carrying this message instead, so - the call stays in one trace; no link is added since the transport is now the - real parent. The transport comes from :func:`_mcp_transport_span_context`, which - is the *current message's* POST rather than whatever request happened to open - the session, so a long-lived session does not glue every message under its - first request. With neither a remote parent nor a transport the returned context - carries no span and the span legitimately starts its own root trace. + With no transport at all the span starts its own root trace, still carrying + the link — the client context is only ever a link, so this event keeps one + shape everywhere. Both returned contexts are built on an explicitly empty + base, so ambient (stale session) state can never leak in, and the span + inherits the transport's sampling decision exactly like every other + request-level span — a client's sampled flag neither forces nor suppresses + recording. Only trace context (``traceparent``/``tracestate``) is extracted, never the client's W3C Baggage: ``params._meta`` is caller-controlled, and the otel @@ -251,13 +257,12 @@ def resolve_mcp_span_context( never fall through to the ambient (stale session) span. """ source: Final = carrier if carrier is not None else _mcp_message_trace_carrier.get() - parent: Final = _PROPAGATOR.extract(dict(source or {}), context=Context()) + propagated: Final = get_current_span(_PROPAGATOR.extract(dict(source or {}), context=Context())) + links: Final = (Link(propagated.get_span_context()),) if is_recordable_span(propagated) else () transport: Final = _mcp_transport_span_context() - if is_recordable_span(get_current_span(parent)): - return parent, (Link(transport),) if transport is not None else () - if transport is not None: - return context_from_span(NonRecordingSpan(transport)), () - return parent, () + if transport is None: + return Context(), links + return context_from_span(NonRecordingSpan(transport), context=Context()), links def is_recordable_span(obj: object) -> bool: diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py index 3c6eb06bc71..57e59dab2d1 100644 --- a/litellm/proxy/_experimental/mcp_server/server.py +++ b/litellm/proxy/_experimental/mcp_server/server.py @@ -246,11 +246,12 @@ def _mcp_meta_trace_carrier(req_ctx: object) -> dict[str, str] | None: """The W3C trace context (``traceparent``/``tracestate``) the MCP client propagated in the request's ``params._meta`` (SEP-414), or ``None``. - When present, per the OTel MCP semconv the MCP span parents to this propagated - context rather than to the HTTP transport (which is recorded as a link instead). - When absent, the span nests under the transport span of the request carrying - this specific message, so a streamable-HTTP session that multiplexes many - messages still does not glue every message under the session's first request; + When present, the MCP span records this propagated context as a span *link*, + never the parent — a remote parent would root the span in a trace whose root + never reaches the gateway's tracing backend. The span itself nests under the + transport span of the request carrying this specific message, so a + streamable-HTTP session that multiplexes many messages still does not glue + every message under the session's first request; see ``resolve_mcp_span_context``. The client's W3C Baggage is deliberately excluded: it is caller-controlled, and the otel baggage processor stamps allowlisted baggage keys (``litellm.team.id``, ``litellm.metadata.*``, diff --git a/tests/test_litellm/integrations/otel/test_otel_v2_logger.py b/tests/test_litellm/integrations/otel/test_otel_v2_logger.py index 455d84c764f..4973bda29e0 100644 --- a/tests/test_litellm/integrations/otel/test_otel_v2_logger.py +++ b/tests/test_litellm/integrations/otel/test_otel_v2_logger.py @@ -778,11 +778,15 @@ def test_mcp_span_roots_without_transport_or_propagated_context( @pytest.mark.parametrize("make_payload, span_name", _MCP_SPAN_CASES) -def test_mcp_span_parents_to_propagated_meta_trace_context(make_payload, span_name): +def test_mcp_span_links_propagated_meta_trace_context_and_nests_under_transport( + make_payload, span_name +): """When the client propagates W3C trace context in the request's - ``params._meta`` (SEP-414), the MCP span parents to it (one distributed trace) - and still links the transport span — never falling through to the - ambient/session span.""" + ``params._meta`` (SEP-414), the MCP span still nests under the gateway's own + transport span — one renderable trace — and records the client's context as a + span *link*. Parenting to the remote context instead would root the span in a + trace whose root span never reaches the gateway's tracing backend, leaving the + span unreachable from the trace view.""" logger, exporter = _logger() transport = logger._emitter.start_span( SpanRole.PROXY_REQUEST, LITELLM_PROXY_REQUEST_SPAN_NAME @@ -801,12 +805,65 @@ def test_mcp_span_parents_to_propagated_meta_trace_context(make_payload, span_na reset_mcp_message_trace_carrier(token) transport.end() span = next(s for s in exporter.get_finished_spans() if s.name == span_name) - assert span.context.trace_id == 0x11111111111111111111111111111111 assert span.parent is not None - assert span.parent.span_id == 0x2222222222222222 - assert [link.context.span_id for link in span.links] == [ - transport.get_span_context().span_id + assert span.parent.span_id == transport.get_span_context().span_id + assert span.context.trace_id == transport.get_span_context().trace_id + assert [link.context.trace_id for link in span.links] == [ + 0x11111111111111111111111111111111 ] + assert [link.context.span_id for link in span.links] == [0x2222222222222222] + + +@pytest.mark.parametrize("make_payload, span_name", _MCP_SPAN_CASES) +def test_mcp_span_without_transport_roots_and_links_propagated_context( + make_payload, span_name +): + """With no transport span at all there is nothing of the gateway's to anchor + to, so the span starts its own root trace — and the client context stays a + span link there too, so the event keeps one shape everywhere.""" + logger, exporter = _logger() + token = set_mcp_message_trace_carrier( + {"traceparent": "00-11111111111111111111111111111111-2222222222222222-01"} + ) + try: + asyncio.run( + logger.async_log_success_event( + {"standard_logging_object": make_payload()}, None, None, None + ) + ) + finally: + reset_mcp_message_trace_carrier(token) + span = next(s for s in exporter.get_finished_spans() if s.name == span_name) + assert span.parent is None + assert span.context.trace_id != 0x11111111111111111111111111111111 + assert [link.context.span_id for link in span.links] == [0x2222222222222222] + + +def test_mcp_span_links_unsampled_client_traceparent(): + """A client traceparent with the sampled flag off ('-00') still yields a valid + remote context, so the link is recorded; the span's own recording follows the + transport's sampling decision, never the client's flag.""" + logger, exporter = _logger() + transport = logger._emitter.start_span( + SpanRole.PROXY_REQUEST, LITELLM_PROXY_REQUEST_SPAN_NAME + ) + set_request_root_span(transport) + token = set_mcp_message_trace_carrier( + {"traceparent": "00-11111111111111111111111111111111-2222222222222222-00"} + ) + try: + asyncio.run( + logger.async_log_success_event( + {"standard_logging_object": _mcp_list_payload()}, None, None, None + ) + ) + finally: + reset_mcp_message_trace_carrier(token) + transport.end() + span = next(s for s in exporter.get_finished_spans() if s.name == "tools/list") + assert span.parent is not None + assert span.parent.span_id == transport.get_span_context().span_id + assert [link.context.span_id for link in span.links] == [0x2222222222222222] @pytest.mark.parametrize("make_payload, span_name", _MCP_SPAN_CASES) @@ -839,8 +896,11 @@ def test_mcp_span_ignores_client_supplied_baggage(make_payload, span_name): reset_mcp_message_trace_carrier(token) transport.end() span = next(s for s in exporter.get_finished_spans() if s.name == span_name) - # Trace context still honored: proves the carrier was processed, not dropped wholesale. - assert span.parent is not None and span.parent.span_id == 0x2222222222222222 + # Trace context still honored (as a link): proves the carrier was processed, + # not dropped wholesale. + assert [link.context.span_id for link in span.links] == [0x2222222222222222] + assert span.parent is not None + assert span.parent.span_id == transport.get_span_context().span_id # Identity is the authenticated payload's team, never the client's spoofed value. assert span.attributes[LiteLLM.TEAM_ID] == "t1" assert "litellm.metadata.user_api_key_user_id" not in span.attributes @@ -888,10 +948,10 @@ def test_mcp_span_malformed_traceparent_nests_under_transport(): assert span.links == () -def test_mcp_span_links_this_messages_transport_when_context_is_propagated(): - """On the semconv path the transport is recorded as a link, and that link must - point at the POST carrying this message too. Reading the stale session anchor - would attribute the tool call to whichever request opened the session.""" +def test_mcp_span_with_propagated_context_nests_under_this_messages_transport(): + """With client context propagated, the span must still anchor to the POST + carrying this message, not the stale session anchor — otherwise the tool call + is attributed to whichever request opened the session.""" logger, exporter = _logger() session_opener = logger._emitter.start_span( SpanRole.PROXY_REQUEST, LITELLM_PROXY_REQUEST_SPAN_NAME @@ -916,10 +976,10 @@ def test_mcp_span_links_this_messages_transport_when_context_is_propagated(): session_opener.end() this_message.end() span = next(s for s in exporter.get_finished_spans() if s.name == "tools/list") - assert span.parent is not None and span.parent.span_id == 0x2222222222222222 - assert [link.context.span_id for link in span.links] == [ - this_message.get_span_context().span_id - ] + assert span.parent is not None + assert span.parent.span_id == this_message.get_span_context().span_id + assert span.context.trace_id == this_message.get_span_context().trace_id + assert [link.context.span_id for link in span.links] == [0x2222222222222222] def test_pre_call_idempotent_keeps_first_span(): diff --git a/tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py b/tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py index cc9b311084e..baa72b5a7fe 100644 --- a/tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py +++ b/tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py @@ -107,32 +107,29 @@ def test_registry_parent_integrity_no_orphans(): def test_registry_hierarchy_shape(): - # MCP roles have no in-process parent: per the MCP semconv they root (or adopt - # the client's propagated _meta context), so they sit alongside PROXY_REQUEST. - assert set(root_roles()) == { - SpanRole.PROXY_REQUEST, - SpanRole.MCP_TOOL_CALL, - SpanRole.MCP_LIST_TOOLS, - } + assert set(root_roles()) == {SpanRole.PROXY_REQUEST} # Guardrails parent to the request span, not the LLM call: a pre-call - # guardrail runs before the LLM call exists, so it's a sibling of it. + # guardrail runs before the LLM call exists, so it's a sibling of it. MCP + # spans nest under the transport span of the request carrying that message. assert set(child_roles(SpanRole.PROXY_REQUEST)) == { SpanRole.LLM_CALL, SpanRole.GUARDRAIL, SpanRole.DB_CALL, SpanRole.SERVICE, + SpanRole.MCP_TOOL_CALL, + SpanRole.MCP_LIST_TOOLS, } assert SPAN_REGISTRY[SpanRole.LLM_CALL].kind is LiteLLMSpanKind.CLIENT # The proxy is an MCP client to the upstream tool server: CLIENT span. Listing # tools is the same client relationship, so it's a CLIENT span too. assert SPAN_REGISTRY[SpanRole.MCP_TOOL_CALL].kind is LiteLLMSpanKind.CLIENT assert SPAN_REGISTRY[SpanRole.MCP_LIST_TOOLS].kind is LiteLLMSpanKind.CLIENT - # MCP spans don't nest under the transport: they link the PROXY_REQUEST span - # instead of parenting to it (OTel GenAI MCP semconv). - assert SPAN_REGISTRY[SpanRole.MCP_TOOL_CALL].parent is None - assert SPAN_REGISTRY[SpanRole.MCP_LIST_TOOLS].parent is None - assert SPAN_REGISTRY[SpanRole.MCP_TOOL_CALL].links is SpanRole.PROXY_REQUEST - assert SPAN_REGISTRY[SpanRole.MCP_LIST_TOOLS].links is SpanRole.PROXY_REQUEST + # MCP spans nest under the transport span of the request carrying that + # message (resolved per message at emit time); a client-propagated context + # becomes a span link to that remote context, which is not a registry role + # (SpanSpec declares no link field at all). + assert SPAN_REGISTRY[SpanRole.MCP_TOOL_CALL].parent is SpanRole.PROXY_REQUEST + assert SPAN_REGISTRY[SpanRole.MCP_LIST_TOOLS].parent is SpanRole.PROXY_REQUEST assert SPAN_REGISTRY[SpanRole.PROXY_REQUEST].kind is LiteLLMSpanKind.SERVER assert SPAN_REGISTRY[SpanRole.GUARDRAIL].parent is SpanRole.PROXY_REQUEST # An outbound datastore call is a CLIENT span; an internal service is INTERNAL. From 10480b8bf0f3ed4f8a22eac9cfa999245fa4651a Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 14:27:41 -0700 Subject: [PATCH 40/44] refactor(dashscope): drop redundant routing comment --- litellm/llms/dashscope/image_generation/transformation.py | 1 - 1 file changed, 1 deletion(-) diff --git a/litellm/llms/dashscope/image_generation/transformation.py b/litellm/llms/dashscope/image_generation/transformation.py index 1347fee897d..e655e2ea87d 100644 --- a/litellm/llms/dashscope/image_generation/transformation.py +++ b/litellm/llms/dashscope/image_generation/transformation.py @@ -46,7 +46,6 @@ else: DEFAULT_API_BASE: Final = "https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation" -# get_llm_provider resolves every dashscope route to the chat/embed base, which cannot serve images CHAT_COMPATIBLE_MODE_PATH: Final = "/compatible-mode/v1" # Maps OpenAI size strings (WxH) to DashScope size strings (W*H) From ff418ffb9c15021fe173d9f0b73040aff17c60cc Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Thu, 27 Aug 2026 14:29:53 -0700 Subject: [PATCH 41/44] test(e2e): serve the vision image from our own fixture The two vision tests pointed at a Wikipedia-hosted cat photo, so every run depended on upload.wikimedia.org staying up and unthrottled. It throttled, and the 429 surfaced as a bedrock APIConnectionError, which reads as a gateway failure rather than what it was. The image is now a fixture in the repo, passed as a data URL. That also puts the two providers on the same bytes: litellm downloads the image itself for bedrock, while openai is handed the link and fetches it from its own servers, so the hosted URL quietly meant the two tests were not testing the same thing. The image was generated for this repo rather than borrowed, so nothing here carries a third-party license. Also drops a stale comment about openai prompt caching that sat above the vision helper; no caching test uses it. --- tests/e2e/llm_translation/fixtures/cat.jpg | Bin 0 -> 31243 bytes .../test_chat_completions_regression_e2e.py | 22 ++++++++++++++---- 2 files changed, 18 insertions(+), 4 deletions(-) create mode 100644 tests/e2e/llm_translation/fixtures/cat.jpg diff --git a/tests/e2e/llm_translation/fixtures/cat.jpg b/tests/e2e/llm_translation/fixtures/cat.jpg new file mode 100644 index 0000000000000000000000000000000000000000..103c370b2e2fae87b2cdde49eef8962077a8e314 GIT binary patch literal 31243 zcmb5V1y~$C+cvy7w79!Vad&BPDDLjg;;=|@C|(wKEADO!6xZVJ6nA$ig`fNJ=X>AZ zGRNk~#3nnFOU^UN>*DJM0R6L+tP}tW3IKq5y8y2%0C503EG!%>EIb?>96UTc0um|` z5+WiJCJM?sR2)oPTpUbnY&=2=Vmy2@0&Hv&IubHUDjFIZTw;2r57dkl)HKw8bprL) z6bTUt0|^O(8Xp^<`u}%$?E+vRLIa=yFi?~LXbdPA45-&W0Lj}Rp`l^^8t*>?96T%n zGz=mX(p#|cVsrmzvY%Qaq>*PmP9)$1J{Q`mG;=w+H;!LjZqu0|ob1#DMYF2%w>%;ND)?+wUzv zLBn8Buw#nCVo`Fam|%-hIS0n&PQ5MzkYV2HV!&VkJ_2sV=+r++NWe<`6)D6)VW4AR z{GXAq%HJ|s0>Z-)!~Y&dZ9V(e6gd|1}JuslP^m_15FxTKqfmqRacM z3x~IX1)BfuKl~%oo!Oee`}i(91G%W&moxL1U7gehG2AmVLg8&khK^lF)R+UWp( zBz+Y?A}Rv@rSEIiJcsiy(ncD2Z1VzzO(+g3bCs~KfV$txPouOg^{TD59+s$=Y}{2d zBW5@4&DvBv@~)7q!P~k;8_)Ud3`xZYufc|A9kIniXK*W<_~K<^$4Z9BD_|H>twt*I z(}>~6?`a!>I1HSBy!fcEGtx@$_tH+me9QX#e2ZKtKSg`0>fXKcVz9NaQyyU1haBM6 zZ7D^%D$p1AxJq@M6f~{&9w|oO;E6;w4(fwG;+cj9qx!@tciCqEUjgMWXKP_;jQPkt zW})Z@BX<6&!ROUR8asQBnm-+hT}!_3bqHw+Rj+J&hFTwIXwYa~yli@{@YH*msLqhU^{hyO6JRGU+-=FsXF-89NnGplr4GkXxdRhDIY{pTYk!^M?X2gMa&u59jWp>v$#%bDPG>yxk=@BCr~YZ;eo5- z?2bWV!BpYe)eub-d^lypB*RgKz?1eoXe*vQ4Z3YA9v<%O&a~3%{-HsXBb1rK6|pre z(|ALrm)+<1oq?~Kud-9oqQboT_NYOlj5R!yQ}dy8JxXVsMoy2zM9BZ+n{RM z0Mu8;!ju$mM78|eG4a<(S!+iIN@09im8Cec>6!OffatG=iGh#dsM%^g4?Rwyg&GCo(JXaq*YMiR~Ob^!oIFLwlb-l^0<1n^Lj6%FO=S z?PxmO!$u{51xt4>?A zZx4(sQIlJ-L9iF8q|(<7Yn{sWCrwYYoeik0I>D;OSlvk3*}A31+2r=#gYFQrJ-BAl z?Xrohe9t|91u|SvJYiE-Fm9%ra9kR`KX&Mt&Dm$jCZ8c*XydoiKFW%?fcT*RWQVro z5Y@gz`jwyQ{Dce^#$MkrTAH((7;EE9RHsuJ%a&9=)a>x&?kyWG=@<*2gZ8H{E9W|y zX}zs3aT$Z9$+6>ATCJ=}r#ZRS%SF!`*YUPEI3F4|el$Ng+Ss%q&XV{(d1$(Khw@MN z;=nOE;0-_Mn$>NBQh7BDO>1CCNZ^DGn zH)C%KdANHV0d|`y-uf}jXzAhQ10S8a=SGA03z<+O*{yhZF?|%9RgCR? z=N^2@1wwSVg@#R~XH940wm@h}le65<3+hAr2*UY=XR^FAV^S&;a6&aH-=;nvbG0^F60#Zot(F-A|gW)qh>vg)Pz$;=G37E`P&9Y30GOUENHtQ$z|fM+5?C)u8= zm`3lwUuu#f)UonO-e>&MiV_Xy?pNbyz1Z^)ZjEirW7h=%Ab9A zq>uCpU@6n)AyHA`Qiv-uVISYuGwfLxe}Stp>H5lbUxbIt8aF{6fFK24->*W-Qp~WK zpl8AFL#*VwRftR$bTG&BNX}IFh7Y@2*&=R&l6R=a%HL^MMtjomZOdFw)G6R&_^ zHr6!PEdK8SL;DW;X0Y*<4+0sZgSuMRd!}(%zZLQ*Z%MWP>2CbPEBXI=+qgcEj+WWu zL_p#vk3T0^=(LpEeFFPv=_=j27>O56Wu8eBwQoiiAFo45%7qG)OZN>&SQ4l4DQC+T z^jxOdRb&KPY7Jd;UNxR!*0%Du+uL&qV~3$Q&1sSo=y9$(%_MDU|3o#xYtxrXu5J4D zbZodKWYP$_&_EL98ATu=mb&D1F$ai{M1u)FiMjruE?LM)eL0z3T|C==jQ)9wq_{<+xCD(_x*>?P-fkW=RW+(i)IJ)0OL{Vc? zQ#*V|2`lhixz5V=UtpM{Y>%=lx*!eFAOS)8w7?a-TV}X!=KdIZCWS zCE}0?X5*Hm1pzXOKXX?d0*5w!Uz+OKEDrP1d18N3B=jxBN_Lyo9oR{!Vv{0ChdP6K zuNo8kFMIAss1SCY0kPZ5_fO0?zrEBPJlzMsb_ZP8VX8_jpdnu$d$bPrAsujJkoj{1 ztTI2M`_iPbyCN$xO|@hb3IKsI-_=$biBDbuku`*eU$3Su*&_(`g+3iu-O+Ru8ed#I z{$eh~aHR&8eU0$~{gSb94=bhnO)qimEp5OdLT{|@BrUC>rM_szksN`~3XFo_-ay~O zdg!Eb11h*Wd8s-3?gXSqinQQ)xyNn*QNOE>kf~k)CHh8v3A3)|z4TNV(pc%TEDq5; z(*sYxE8Yo2@|Ep&mtN!6zi6Xm;;#>u94S=#9dQbI{{BprLwzI9&i7f5MzCVzYSSGx znsBA$rm-_khpocTM$Ep}*2t^&cNX6CFsNn;LQUsYNawpM;?pz9uC(V8O+ z1U*k;S5NMPH9I@kSm(0bbxo@3aHX%1b!d@%yI_x9cPklevY26@4GecX1+X_EPrgtJ z-jJ=SK4hR(74}glJ@2W_juOD^)_ChZM!zZC7U3^CE(Lu>dEVvp9K4J$yC{_NfOGFy z$I775ZB*(g?9DrM1CbPgApksAbKQ}MS9RH!>yU3gH2K`^5@ z%YLh!=nYzPeh&`@j2R|OOdJ;NSthwn!73}&$NCR6VQRb81JrE=G^2Qg*3$|D`~Iz| zU49f!P4rSH49zHf3KGrVI@yw;^?6tU7AK}8f?V4K6l(~@5c z$9pY42^d?daUWf3!>2S}0k>#4zDyWgTB-x1gUZB;RdiRelSU6 zsBfAIvV%bibw@Ltr+fwS5`Kv>;jASV_P9i0eI$%g1#UlT3~mp1@ci&Xw4`A<_Tr{v zphCqQP+ITtPH*4m6FtO<>u4T0eZ%Yk2EPJ0nZ^P-_I+Age>JJjt^c|G3 zmEH3t5Lcpt<)RUrDtVq3J!oTc|bJj^r674gP+|==#{1(V~d2`xCV%sOj!=PNDdh#Qf8u$V%8%ZgHKg zy8=WY2!O~7N5dYnX}+LZc@grSOd3EtY@?K%-hXyps2*cfJ*~|cxt%o zE{Bo@FY$<;I-XXY_d`jOY_J=oiXz zK`+Uv`c3^fe^8A{yfo2dL3wIeJSiV}`0VzayU%}VXy9sF*d&updgUN0{IPBzGC$mN zuYJ7to4Q-CIIUdR`%~YaS2WU>!r~tex4wf_A1$56Du+6bjNT`_U=^Cv1oWyAIC;UM zU#>>Ci(NP3RBbo4pMmIKXdfKaWV=9~(FXF3AeC>AXK+<^KEh~ayZx&l$lXAZnj>UcliYgS=DBm~VX{|vgsV<7xewam3CK_&be!7NU=8WG%ikw!=kqSmlQdPtBqr@(&V`H25bTyfkx=4YcGA?{#e zUcjE#kX#lq_6^vWmCgn6M*S5KvUqEN)_18D#kspvffh0Yz6<@F_({>{*0{OXb* zlH_vwE;!xRim_sV)jfTbe)nuL-74Nf3AMajmF6PbUW|hWU z)h*;g%gsK%hwx^g@4TL8@Md>^MzoqKnP7OhJTmr;lCPqwV~S`o@F}Vy+keb-c2%qY zvL3OwW+3kCX8aI;mGdv1k8_J|22DC6+_OOCVpRHq(P1<&A8!QxB~N{rv$Y;fH|4uz z!`}~%X8s_oo>dhCn1U|$sh?#)FK1o0fjIBD=E=v>Z)!0+1DB@Dmclg>=UB_{)8V(C z9H)mn2k>3iBG zC3bA9vQ}lMX#OlIE2DCY)7c|}?wNFr=_t<+bJpe0JR(e`DvBq7B5>TkYFRj|<%@~d z3)A#zhs(Y?-fim0IQWsFP-!oW7&*_gEsZ_LvL{N6T4WaP!Cr~kH|c1+a)!1tr|an* zRos<%aA3QF4Tyj8xsP7>IPAFNwqm5A>yrpM6|}8IC9DFMSRZ~^D+pRo)i*lJPaPF& zaJPSWSothqIY2M8j3$%6w=(OxS-k_ zYTk2+r6Grf%o(r0!31~i>|8ekHg*0Z=&{1R1#}d+sduL9vzirZ4!d|@TQqQwsB2E6 zZST6cSGZk=Or+@d>(Vy)onu;s^tyYW1|aW4 z12>nxe8jXgwH(c%q@<-J847RW{5R!^82B$g7jx(@=Qc~_O{6dn{=xYE+mikxmB&|= zsCjWmp(x?s7kK(Yb9FjDY#`>GO}$DNT7LAaYft)p1nhZna(TiO8on79MxBYKUQ)b4 z;mK^m@oC*z7oP|{u`{kZ&-cb<*F~u7wa+SBwWAAt4`$(*ipDE(uDLBOr1J(ii5h(7 zA*U6owS5>QBiVkfXbM_UKzZLJS0ZE^tR^RDK?~tBCJ)OB`-cMxp`F9 zn~b9IC9))%jeu)F;g)-!2jDWXp zAQ`%>;{v>; z$S%@%|7`n3o~LD|m3@1EbUuK2DZpTDV$Qnby{_q$koV5u6;O9n!Rxk1E3kve6jO24 zw2C2}5Obt1{yW0wtIbks!!Z&BP{KFUn?;Zh1jNm3uePH2C`TEvOVJMQUtaA)z0lLr7j-|JJs(yw>>pavS}=93iGlH!fCa9bF?*D@D6eYSsqx`5 z?V{*0SJKH1T?f4KOS;qj=A5gKbS{+4c3c1SQ+56l%D>sSe@bfLU&~@lf0s{myT3_i zmhxMrf7?Y{|794m_^_~z9-BcP+x>@p>$W-+$8~S!DtANUhr>S3tyKkPZg#K8oEL)% z`)4bUFUN;XWtn%EWuqEviG*5Hb@$HcCJoMlN*<>-tqq$XhwAW1%Wf$9-fwy}sORn- z_fRSqWnYw1l4fL?LgGW0$^K}O{o#8Bq}~gn39~XN$Z4OPMaT7>nFP_H%747adRU*) z`&02_q!(ntO$(sQ$xuKWm-D`+fvw4*Juv_HhMVqbWH(ZWihL4pu|z5#zpBPru~!L7 zOf!*Dyl4UjqGbh3o(##OaDt>1w-ux@P<(Qi&3rw zkDtPKf?qq^5OrRsU(mAAmr6Lb4@2o)T~EvFb^1xm^6Y47Z1{y^{^_=E1YDYGAxI`n z!^LDXf**<^Dj2t?ff?F{-m@29#J|pj&dWr8nAI0xwA&H<9l+RuG;otNZl}>OVG(Za zY#BCjR8b6m1<2@)$lIBbR%U;j+Ju;GkIz^Q!-z{Z*Idu3KUk8|Y_>ivhs{I=3W4KB zU0vFE(i>tdAoRHfRbHizEK+2s$&;ulI_OhV!;j+IxX!ED<(kF1A%ZcsFGT%pO0gQs z3b0_>=LhlhV_PBdYKu7K1=ILFcY?549<^shAd|lIChzjy?~sgeQ*A{I#GpKx0Eg(# z!=rMI9~wf%DWWiaju14;d117y=p=q2v&O`XFG>x@50Y!x*f$_dK7eTrFjFenfcpT= zuwR{Bl!aB(q#O&X1kHN#9+o8L$-t2gC0Q9hys$)I`8)S2OyEEyn9?2MX2x%n>IZ zmE((0n6o3#K1;z-Zq5kq84pTvGcVIE+TmN6P5WsIo81kQvP=owQ+)6QF1xPWHKVBc z-aw!Sg(GVSer+G0MXZc$eB4>+&T1@bj>W!^h;E803RqJX$}X5XdiAi$EO%yc&0!*h zJ1+sRyf^)xKS7EHfmS8e#@h|@2lqU}-bg-&Qe=MiLwGwa#oq)C{jZhYe^EgH?sw7A-tah~oA7Vrm+?C9?G|sJna{j)lm!x^b|;i)F1K6MzgEYW zdDJJpr?4-~3K|W``2!hrar5fFAGPS`JP*6!U(?Ym4yB7m151B&ufX9R>UpMxd+FDiKH0qe)_uzng zx~iZf$ywx}Op6xxkiN5Hzm(Ix@~de>)&?Jg$W4`30JKLP1pzb6r;lU2s_M@+eXNCc z-^#tI-HbD$mbQ)Zz6~7cuJT7uw%TZ~+={UKzc~dCkzp^Osl{YI%bCI|<~xI@`S6V= zr*3WpiFP2QT|Nk-MAq(vt$VR+$3Q*(A>>ggpC*Hf{77V$e^Cc{pf7G_RfbH|SFniI;!^I~bob{<8JZe~ ztWv@_MJ;;u{L`2|bbP0zL{R@PcK07!LL@#ZzM0hjl;en_xBFnDXUDNu0F-93%-%;% z0CQvigqXX?m!#cQ($;qjS_pEpnI?QJ36Da{Yffkrw3J}$f$KF1Q5ohtX9-a zX0s4l1Ntv)$kdMlIlJS}U8~KGcOWy)SV+?!&&jS*;_Xy z@=psfFs3%J!R=$u;j4Grjye^Ms#6T=jT&wsr^cy<8a4~ySYSs%PVu=T z8c~lbM{qqAxqqzQWyR$F74R;(ptp3{#FKIdbL-le>m?&8dj^ZK91H!ZB9yRo8PW6= zP&Cl0g;Hg2iKd6SeNu3^4bH%%{!D-)hp)Tk=9selGLpx`UR@CPod7wlOAsol>&pv9 zak>bv0!)Vg=As4c(M)CK?jLp9pOOZt+phrc=?R!E<7`So-AlXpAQwK2#&h7)he#26 zUtyAu6H>O?Xv{A}5X0?$`*89gk(4DFby2RK&*>f!7M5m7eX9Z7CTFO5YiM;LRh#z; zLbugku}aRYUm;NeqE1A=?mO=pRIKm5&HGXUFGH2DKjhlQvwdXLM~h@Rh5TC3Ui3$s zGo39M6{q3YvrE_M3hZ<(da+Pc+tk)h$>6HaY*{Uh(#uy8Q+!%ED~y&DA-qtnJ3H~Mk|4WmtcHC_F9P!0i$)TK~8pjvIM*{=Kcwd^4 z&}DgH65Mn>zs8cg+fNDB>ZJw;X$kSyw(Op|=j~QO-__EM_iNccq~FY8&z?HwE<1JXR&4(>A@Z?TEIcsRV$+S?|mfnII2=zhg3CQHX>NQj!6p;fu$#q*7?`j z%Dl4NeS_%ou7&fYkz-`^u?!T$sf20t>OaiR2r{2h?J}onJ)iND#!E{M_cYqy=(WkB=O>|`QXs%$nx{6hL3G+|^ zE6R9$S>Kawokm8j!ad*(T$K8GF2KqX~eUHoRDpm9+S$8*LYvfcG zY@1cwzwM?$V8w#_tiOT>4eE4koZTOvu1kx5SF%j&8HAfUQtjAx`onM;Sjwn0^mM`^ z4wL7dW6Gou(bTbgq`2QY;)rW1CMO_E>;9>mF`Xm0E$e^>K`qLpVOrsBptr2qnN}xC z^m4j-x*}Mxtm?J)muYG?PjHY{S9dK1r+IQS$rK3oc&YSgnTZ(rdb z8YmxWh1FQ3#t)2cxh#!87qlr+`9%t1ZrN%_eRo9+6*GL$U(uR@`-f=21CnJlv- z`%x=^Fs;r%4p!Ek1X&+3`+{GZHGL3*%WL126jUYp^ozS(Q{!PmJulmC+TY*Soh*F|>1YnG1Rx!$zYNrRP2jazsh!i(R+z9kW zfr?U#N5sE^^$r{FblDbEiW+LY(r;EBNP{FV`1rpEJGmQa7XDp{Jaqg|ZSt2gft7%r`WuJ* zZ-*eeWJt&P~;=H)GNQ4C0;C>v0t+0qf_T9}+9vOe>gMc`Xs)CitQIh)Ixk&F z!8)sf?CNo9&PT~zU5eJaiSo@dWHCC+Z*C|8 zLdBl<7HSCXkfvXD9+`$^50CP^qk3sTc|d9&W8Re(^7mQIjq@t;{9CSRDx}GgC)UQ~ zz|I^!6sV_#UAF8zZI;5Kyb5!b-nT^pZ5d;0*Gd<-Mu9U~fj@XZZK~Cbr-@mnXMmh? z=X;1Vf09-5;|X-0E!$7bGTg1_u}Z149D!beD;2K!$-3X4_oR+lulC5H&Ky1|rzLJt z;7Q|%U+Si%pL|Qa>UKUZrBYxv#&sZprHu=6PY;t$$iTN)okp|2wOVDewNCqdB=d`K z^~5!hPL6*Z@2=)IQ>441b6E=+R*$@CZ4&v>3;Etw9<^8V`bk>OCL#6~ebc zM(K!R1p#lbCnirmJS4ksrQo!G{kmDJwO+|DU&GdFAv?2yO3V9tEEFn3In+#@J* zw~N+4c#;-AY|}{AWaP=T2&trqgyHp2i{?W-+_uB$7`z!dM%xLm#3KY2?IHD7facvm^A zZB4|x&tf2m?8(4Ax}ZxAotAXA)}rcwJR9*RU`qR?bdX)8rra~^F8|!C^bm`66!BB1 zu73aoTcPv7Kk=<7K7A?0RZ*T4y;l7^D6yZp9H6sDvOF&ml?Z9b;bazOrL^{8rbd^6 zz!eCZbt*Q^G(xs44E87nbj!X7SgBnFq-DsMOf7{Bq~IYad*jG%F(8?L<7*G8XMuB!xHt;4Jd_?)H-I|2a;hLpv}u{>pqM--BQ(__=|`88ROp_Q)AXQR^HD{b z7{vi=@T#5j7qr#XuhSwOaVfxT)=}^s{2!Uc%ETXWfo2-`Bo=t+Q#Jj<5auVq3gS#Px?p?o%-S`B^~A?GzrTsUNDyG4=tOsLKQ&TSdHx4KPUG% z59t$KD?t6C8-89-MIjWtXP|;_fB)@`pv}BmZ6X`v^ruTnW5Vl6@nyC&II`6`u^VHN*zAWb$XaP5T zR69I-5$LNhIPQ!uGVsWlZ*Hy)OhezX(A6!(491*SFKM&Z?nAL=9>fMVbYF1x%FcFR4J~mHT)(s;M4mJ1|X( z);#gN9MhrgDp8n(gwj*uFtJRnEXK6?9oKi7`x759mi_8x%3)=_T*2D@H7>|UCV2j9 zSYA|Ega+T7Dbh+xe;~nq_*iFDxQhEdeB?XJ8F@0QZ7|X zh6`6Ch7915l64W%@jxYvwgVCBKhhKPhLF51q^I{Yv?Otms_j10vDIF3c~IUTC|OWG z)25(s2rqpNSpNXa7t7u39WB}RrL^xjT+F|OkA6Qpu4YRE!T+hsni+n-XM^hqbbX_v zHn*8-!R6E|1ooDVxHy;(Tf6}^d>j?L#ZzAFpgc;Wjzo~MB_<(y)Z86iTm0kh-aYDs zhe!Uk7viI?R%fS@!}3SY#I?cx9)2UR+5mb^X2F(K@R0HMZmgO}I{BSu9WO8Es+{C2 z0A<2p@U}+d`EvwBooz(u8%KvYqY*_5Fmswixb46dYGHX86q+3bBc}jELRGk%k{qi$ z&7ym}+-EJ?(FU3)3a8J}id0q6032+MdBC@QXil58sr=D;(}BaFT`&M8N!7dM6H99B z5qamnalTM;P~_CINN>al`#hlYm_cL@ydgn@+$jr5Ei`@h{(RpE_OuMz{l1Fsal9az z7Oa84j&GOP#Y>Mlv;vRSrVp6@O>cA=A(b^sr^A`UEy*p{XP1gZDE>pb_-mZ$$C1kR z3$jQO1wFN1?rLw><~@2skTr~SfXA=FrgF8Ez7Km53Al(uJ`QYQ;pbHb@IWKE1{7X6 z3uDiUMG?;6?Zi{3os1!3DXaI3v2Hx8y2GB1<@TdVt2W0C+7BQ%z9d+j#B;PrON7SW z5(;xx(Qp66JpUV8p;+>l+KJjiD>eTYfALp+(f$qG*iYnVg4JYZ(Zxjr$EKcpK9zQ3 ztU$>?R3}`7pWaBzK|lO2&Cp76w@007e#U<;{%kqZ>x5Ku`-fNOMc#qJonSWqC<}JZ zV)lwV4;R|iR(GnDt{KuIj0@+8H@j!&#*{49B~EEVbJcp!;ymAbq;z@(jZrV0+Y4llHpi z+QF#pZ_T<&sBDFSgA7(Z4t(x+{s474`5)n~piTNYgNifSJ zG<&Up?W~ZQ7Lh_cLcq5o4;yCzdmdWFo(%L%}je1jP*)n3e z95oeWW@NI6%2hP|1(7I5>^|Fz2O9Pv{yb?(d`G!9?%P*@(ovhOd<*VOXLxO6+J#wr-S#x<0-Rfn?9PQK8_Q?7=b4-0jp&ICj58?kQ9;&UD%3<5}LfTZ~O*#_ZL z*Ql?!?(-xS3ga25W$k^9m?k%txsXE95?7b_!Nox;*_xUaHoC*3ib1eIxY+$&MrGAD zJ{9L11I9rO)PLBf0h_JVFP^sPlkjT0I%iYpZ;}9aooR~30Wd}1o8=2gQ*MMW`8%(G&r*e8FYu&LBp^5J$7)m4Fy2M5v0OHNcD)@k?o8NF%?D`I^O&)`= zs6qT$&)S8Q^9aS0o`G{7s$2e?IDGL!vzjARF#38JL=TD!BoRwa<*AY?!oqyXw2)@i ziKpaR+!JOh+LU&fD^)yk(bTxr4@=ZtfQGDq*j&BE>zZ-YcwY7y=)A3 zl`Jk#EukAvQ#((`d;0yL7`$X1i;%#xix>?1d$YPI#}bBgx2Cbv3%u}GfY-$=ME&Yl zQIH(Gpr2Z!4XIv?Mg2mm2)Z*BpW`>DW<`1R#fpy~J}q{<;CWXoRmgrD0NiKi-?W>iWT>5YyI`GVj+E>7A z=Qfipe8yu9e50I|arn30l2VH`t@%6!zIc&cO_-UX3_XK|wq_p11216a=jmE^Rr*sM z0H`>mNJ!T01+@S@wT`OhYa%ZrkiE}naU{EJyeN6Sr~q&g`LT4*4nixIdcB8@%$zmC6M#|m)l ztImqScIv6;jDdo}rR2?WXhdWT(}L3Z3Se!HkE$$oJv1pzx_6rM#%z^oi?y=Ky$1ddh!CnQw4`@812c6&8sBGZPf~u{;Zw&>!s6<>iv}28 zOOmL_93#zYgxJvumW|KS_4E|@V#Svgpxg*iV#xF2zml66eJWkqtpBt!BU)^LQ_D_F zQSER-hn0eu%#dEe0cL|qop0IT$&ChuUpc{&RV7a;_Fp0(ZZU3hsHPg?BRhYLyB{7lEXuklo|S22)P8#{kcgNqQjxMykx z03U}h5UEz?^sQ<0D=E`SDT6F!7A!c0tN^E0UT$md^SxSfD3U0?%H=-4D5&zaE&^vI8HC+ zAV+%#yUpprlHbhOy9$N!72fhJ$$Pvh5D3?zz`mFuPD0S|3@43G6*3Rd=oPTye$R7# zYJsMqc8$4%0>6x4IkW1ueo|Z`PVLl|tV=2)*qLDLHO{cbr z7rBmoe4Ya(N*{@)z*;H=mMN{*Y*VU+`wBiD$`dwA>1!6r)5@4txB*8{pAiiWDv=6u zO!Sy1d(^tJ1F=w`VatPSKfAL((&Uzc)Bf(Ry0mxdt^*k1_TKPwpN-)-QS_OUt6cis zFTvdfC#k_)LhGjDE>b#otZ2Wz5^pd3dj7a5J*F{|SHSH0lx?rCVxmRW;};UlgU0&* z2?PJl5dD=3tb5a9u%9IUb{^nwtHi(wuu`Z9T$x+FnuOdUzSXQ{r|6m;_jiArGa3*V zz8?mPTpp`)(sTV50Od!b<4ohJ+~)9;EyT~Hc0xl1=89_wmN_w@ur<){nCFX+#4JJ+ zt;Dq5i+;;BnJVC^Qk!wJmm)Hc>BMZ@Znv*x(A3o$K3(5yu^gUcU>5VLAakhe4YSDK z8(YtBSSyt-c@fl1fWF{gNjO^sW2epZ$F}O2yPKtT0R*#nd75Jn%0ZrPgP4BQJ&QYF z+?z}#^DU{|kc z`GZSd;_4^YP2t;yC-i+TqO*}`kzkGRha-5}jnPwQCNqU*y!KasQQk!H2v*ciefq<) z8=>L&tZ(;?{@1vV=t1-Z-h7r$lTlcXlR2Qb`7DcY_F;8H-YKQM;0}n?B2vHrl85#* z4oHRwNI~lL@c8gBr4v5!T*H(#?1H4o4CKgB!EbDFs=yxY>8*3kjY>l;HzZ~BBU31k zP;KV~RRb&4=mCOMLun7qYiLA8!x)M~ zFqNDK@74MK^za&1#u%uso{!n?Yj(c(6IafsJYZI++B8H!)#fVxIAI(5QF{fospb4E zpL1B{-Sh0_=559PEdTV-Crq?|>?sG-fs<>Ol;21_AH8PIt;ZB@@^a347<%#6%j5(`^ITW)1<7V(i`UFYJQo2^M{hF zR2vB=YQA4zWoiwu$u7-f#$)qY6wUz&P9j0lF3!*vj5i;k4Ml5-sxh?a*97 zxfyc96eL@jk{@c;T5|bvw>htWbJahQOhJvJshZV#=*kYnvp%6z_A^=6P8zv-e017~ z%?q!#c33jTpp}ubgGbDM7sb++j9HXAiZ6ECzFnglDjB|s!E48R2bRA(hJ^sN2_}~) z3P%%2`sm>GjoGWk@jz%U3L$XPl#Qz&CxTyYToibQmUK(PQaOTXW>oCs(7QE4jDwLO z9Q4VIgX7LSqedXm{S0-1kpA6*xMVL=v7LkYG`ay%PKKE~1XH-Nc}{dzvqPUvnhIGl zxwKY5uyM?0xxWda-?5L`Y)3uxqdFN6iL|^|W}QUEWGl1%gW`7w6!4~H$klnjRdgoL zFV`y4-bCE@LA>?V)e629HBOOFnK-G2%(84_l3XtM2BuI5Mxti-s`%~v!AOxD)U`G? z{tS1`@++XpdrOb;SM#iwe;eLG zI{s*?miW+CJ5Y4t3k4i4!Vn8$q&zCi@e#W1x4A#oiyA!onu@{@l_B53;h& zqg)!fq7v%RSVzm%b;X^Rlq)@+D7=Dd4h_{fDszvhSAfmO#lWNTOog7J;=}{EdQ9nB z%D1EQBwd^n;@^-Nq<(bRz^syeM}hA`t6Xs0PVX>Wtxl$2C>rC66MUb;KZf;x_4Srf zZARPLHk1OTK}uWP36!G6UD{HhSh3>4oe&^|04)?K?i6=-cXxtIDQ=-aa3{Exew=f@ zea_zRd-nV987t%Yk&%(L=A8F^UyFJ_^9XD!S$|vJBKLCYwC1WT3B3WO3N3^X60q%Ej;a5pSQ2Qqwmfniw~?^3F)NANlkYKs}yUReiinP*{uU z?t&2gz%tqLoUPJSZ6Wcakgl_-%aRAv}7fupdRta`X$lBbMSbBa(hMmMt_8` znca))<*2|gXr_;~O0SS6X%rj;M7SFnu@@m(BcmS(6p?7R0SwZIu1sK7IbM=$O*drN zAW9;Ssc#L3DJ-KhbSyND&%q&DXw(g)QQRyHPi0US3TP+n30tX z`=jx_IcN-ryo3DY3&18zu;&?S+hm0KHy3!WuUa;mDr%#={wK}5wu>9Kbe>?Z9=ApM2zKtavZ2Msow~kU z+Kd(LQrq?n>`p>DJ~~A$9daNT?`WeCcuG)j5ImEa;k1xZ6Tm-s{B(rk#EUlCN}(_2 z3I!cU{9d``>}gFpUZZoxo;FF0k>|1EOv(f`1Dg&xJF3qi%pifO@z0Y3^Nev>2QcL> zZ2Mk^*A{JXuJsRE|Jp?|kN=bR^}k%O|4fMak5Rm?-%H*B!x;-3A@)8G8>5#<0|EWf z&eRwN8o>+0s-pfmn;2kN1QfTtVP~_(s(}YkCHLY@ zD@|nWS^#y~wP`oCRQhG*$>ZA;&^CrPU(hehRG!}Q?*}%V)|`+Ok8m1!CNFMJ34tbgGJ>{qrazutcA^o~cKqRmj@BV^!Oqr{`mmYxIqX1w*@dbjr7yf|=blB{b;^v0QliLJ>8! zoD~2gFTQ}e#%%~9 zL|Rt*!1!8v|LX>m{sLk*cZAWl$q@CV2*&W}R!IG*tlj$(baqsYi$_BulV-C38s5oj z2)ZdIqWR@{T#o5Vov#FFnu{78(*r3TF z*C}^TtrwL)Z4vdM#6Q4U?xt1fsB=8sporrn)7&r_RMaoxm)cwdco>$@sTgp^!heET-i}PH^Qpn@)rLLBLx#!6A8w>h~Y&G`nj~U$Hk-HA>7BeXu{9AF={lD zvX)}h1xKr@C@ZHU9!kF%ny}Tgx-%{k;(FfAemo`|B?QVwLT_b%R3dm*y5J{MX(X{- z9keaB4>dkyN4^>Tgnl2+xc4+6urLoNyO|s`&-h}%$Q`sRPxHi8wXT5`Y-rt!??0)a$&)`L=_&@+TGjW#qmxW>Niui zTwkrlqP=7o(sVgY!Vn@8Tce3&0cq!?cNGp$XnD?<)U!fzS3-HY=#RgtbuNWug)#{XAKwkEZ$%$8$NSI6*x9^P=_u~6E zSyYkLV+XDU4p=MYnI(}Vh5q0D(m^NBH!&fYz^p~p%Bqls2==~{@jx8ht0XFrB8F5@ zgF(Ozc8Vk`f9=oCZ7POBrd|%C;7ngAIn;VPToe`1Jd3si`;d6Sb<@XdTT{RFvZQUcDaBplqsAN}^i6waESd)!C^I*KnL z^n42%#-g|rlyC>ne`&$`epRyz4GCI}jD`vPv>8lsBq)~+xLkyLT#6$GBJnL4 zUb*bFHfw+tGOb}^-%B;>Vlp(-L-1}};wFxcK+}MI)8wAYtQ9IdinRgMj=7KDn{#`s zb_2C`{R!qQgx(0w-c7Qb7e?f}HjTk#P6TqJIVU?I-#$7TID@p`&s+E;+l`BCCJHz2 zx25XHL7XDt1zWwluTjKX*yss~u&O(*c*O49*-14byC+Rq*ze;usxXLGG3~MMam((^ zABCgq60UDH0Y^Oy0acYKa`(h|=+VTKqcXy2NX0KE|VxARv^$eblz8WMn! z1`9_DX}5vtFZYa1rqB#yK{eqGCFo~fmXC)zBI)m{hHz%ZK<9d=LBjQRYGcH0Rh7-x zf%hpm$zRgiBs3e*U)bBu=4zn!G@m*bV6-&|72>h9pYTH>lPbyK1G*P{TdZul+SIl7 zMIM;BDdM;ov)>Lpv-cF`gFaqon`1GU(O%5$ql6^Gt+B?FRQyfS5-sTv*I#oH2%C4;z)1^mAno@J|~DwWZUqcF^wCq&TjeNJ5gtno*_kq~CjY2zkt&(7=ap zJ|Z;j^7)fx9?S+fq8t`GA;g|amE)@-#~&`UQ;gepIZ0Y~5W9V0Ptc-s z(0GcakY!##ryQ_ymO2zrj~CwO!u^bJei1^=2@cqUKrEDk=Q2plpp-uq6Gw&asS*hf zmJN&aZ)by`PlF8j*32UrMn?eDn<>6`4UBy^UrWLS$h~k^3C!oIB*IKeg@&1HA5)R0 zJbQ;*e#_c()qhSAK8E)1flVG?kPcBx8iFdoNkA;8rtTH*#cewk6FNP>s*{jM>#(QAtgfd$7wl40k#AFFN5;GORtCYAEuef^4Q<{4XAi zGr*IWndOb?IFt$A#{Mp=zYLSQx2&bFS);*FN2jJk*(yGp4JHws?~Uh$7WOuR{9mVe z|BrA-+A}gl^-Z8!F##TWwW9b3POZ~gB^R*b>`ag4c?6ff-O9&GYf%ib49SuSFRD|; zchr{jjNo!bnb7chLs(Cd@8+0Ysa5D}%7=LTm-)sPdr8~i?BeuL{;Yynm7uxhH(tzN z3RZ47Y$=1y3Bf6rrv$AZcHbV9`N#1JGbIjUfT>ZFSxt}bwQ-m9?kg z4*PDoCAnsKL95U*w%F*>wyS!T&^<^wGfIUSq20hxHyp5BVZ^M~clErSWq`rP6*Wf# zsW{iaDK@_C!i$OSf3l!hzK01W;&V@5PfJik^r~|362bz5;~(7E*tMsEF6|+S*JPy! zuH$7nDqQl)5(KJ03D!+;9+F?g_q(IrscNt2G*#F)94)+ERqbSswx$iQW>6GZ z(TyV$qT!u?9tFIpPEVVsDfG-=bgTIDG#CM<)AS$bZ%tzO60rp`3m3$3QwyB+N}Me3Wi5F}qW{HXNJ$bq?D7i?JabP)~+` z;GO8yay$eOXW4xMtQgm@sIxF7w3J8A4h-t@@NPhS0}g%=27(u>cFc1T;ShG259P_M z9aMM(@#$}mdM2wd=59z%Ui^%HD$6}I+exl-Fu-zU{^j?Gq}j-uSizngX6X~>!+n3H7faSMaBjm1B| ztNLh=(k(#1WdEdzj7A0idL~u<9Y^}>;#6@vNRz`Z2S3i`?U7S1jd~=l7k*6D(+xz+qDYm^}dpOpCU@_?yhXe zs@@WL^;UG@#)q(8g4E8Kkg71JUuoIfx7@mJ)o;4?>(5{3$s zwfn_1H@#74lcv64WZd)lM{?$$OG%dhEII!%K=l9bd{5nHsU8CVeTs!y9ZQ-5(8{(jvs2Rg zmA$$7%=BjOcER@Vb&hHXE^TjW+BHY)fk%G^L@Qz&Ny%hnPKc)wcdn`B)>Ge?m21y* zYt&ZH2*%t64=I+wkpzP$Le*n28WRgD2T53rB?VyA}R~8 zuno0+bONB)*4V!ZK$F5sTA!S;omout=1t;S$puY)btP1eOG(|*`&gb6DnYgWt<0>t zTuVF}yDxs+{`W>?-s$QFwTk<97zk~w?;fq<={pIpRjQIvWllSCUgJw18~S)xhNa9U z`5SC-#~R6qCZnONLE^RwCf_;YDfd!5v~8nzUSVE@ zi=aAara@~#c9*klAqQOL;Sy~rM&x{Gt_&vVIM2rTu^E#EPM9PC9&sX{_8m(RfI`nlil;P7Rjys+# z$|WXdvx@aqVMWC{B*-2d%bUN`0zp4dQhQ;ON1mh2s~pC zS<762;S!aBl)^{`i&_dy0#O$<)_#r+8)gr$#CltpanK^VZL~?sk<6HSW)$GGgcj+H z$$WfYk2?*`4iH|4DCt+$$mz>rqB`Oba+wCT#~(bC=HmFHwbjg2y*DR*4h0GrQQtud40d%FOdFUqYkW08)nzPJ!%=Zy~&RS#>$Q&4rr)tPOXTz`Ie^eBJ$D2awr8F=2c zXU_s=j%C2{ZZEHQRy3JGL*cR;n4JD`;Wp(8vqk#WXI4ja3GYO`EP$nfPtUacw@x|? zV-Mp%POT%6jH)C{srZ~vBy)nhE^Gjaj>d~vPKtO?IqdZ0l;ilF6YU_`cJ|Tr+2d`U z8NHR1y95KG&g_i&{^Rwmi?7}zN59-_KXUFk+5N`#l{Eru^@`_S2*EJI1~kv$6LJZ# zsen37e*?nu1k+@HN3YrvEu~7mlmEfpi$$p?h|~QS1J2gYubcU%WWwxU!t?(VqyOh8 zuK%nJ|8uYNfWi2E3)!{rKCgCZDA}4>h3Bt-FNwUxuKjE;^q!rSc0!1>xq?%C?>`ug ziw>_$YkR~PNGpYCT1u0tweb0f{QdezTNds8>;+H%VhCs%@~z+-GG*uAv>+) z+M7Fq@EnUOsn9Yu7u6BjCHFCJ1Iqai-_dDCb+w8FzLS5&%t?%IVon^AfQMIK&V(TV ztWYdck`4k5<`p1-V5~$ZQC>ogSb*yec(F*WV>eIADJc)}lS0yU0O^gZ+pl=>NBe2( z`{2>id!Rv7ZnN0-#etV$x{UPnJMEy(p9V@}=pUPDPCi&cz_T;J{+$pGos5yXWlZ8# zkXIJfV#O_6<=&vZ@=g&?@uZrVHx&9K!BT8vn9AZRl=?^)obsTN{&Zq)KHrJkyo?4loKVT9(l9TuY zCu~@(n}J*#876==A3EN0#GZM)^)d&HKfydKtgFLK9%!0Ns^Zq70mf#N?oWf1Ed)n~|3%0UCx3tc@Jujd3_TZ_Y7wJ_kue5SnN5 zX0u>1J&s&$)j68)OmUt!iX`1qsE7g8kG!<^NZrx9N@MaHBsAXKq@+#v6V_Qg!0U5E zSP*0uz(7F9P3oi|_yIT#7g>x$6S3xFy`5Q_*7K~h3s|~_r1e)B9q}QbCx_*~8Ttqb zO>+c{b_zDJ)Q7kh1X@r0{w8XST}Lpq-!12JyJsoTy>$Zl(p`I8`MVP*O3 zw){e^<}yt>$0F$^fqyB>`%@>mvg+k&3*#t$b$KPiyz31r9Sg``4D}krJASRwUF$^^ zY-&!ddqb_Z1R9Jgy6mX@Tr)kiE1{i7_M2rvjT`LH50e9Czp+|HjYd-Lzbgs~o)n>M zINev>?S;6kv2%7w9XKn{W)r-!D0&N0XHG)N58FpphW910&lFV*FZ({g>aEOeB&pZ3 z_Nf;mjdS(-Me`}vtQq<XaaXZ1m0Qn4n}ulbsEFtGJ&T7JtVF+( zr(0f|3_q6=319U7q}?zOku$)du*K`0$QPMPyiQafiKUgH>U<^XrlfCbpU$ip9cNi;&a)Fn;Z}n%8qSVGqzRL!FtX%jp>=o$GnW zzV+wEar@5R>)~DHHq(Kb>`&ujre>7YPiADd608+L814N<>N@^>gcLe&?drcDINpgg zF#b9Ec3X<>UG;*0xvB1K1laBuDBo5z8gSHU{(9ft7uZxrQBPS855vh@w-jtE&vpBh z_}Q#>1ncebT7`XvMa8gcjy|Mbpf;L!<0o<7%m9Uz=xSTDvmk&dsIUlHDk)o6IXG&T zMim~8u>0m<)&T!m7jZ4kZfF+uaK4YDL_;q-F4#JJOa)@$_Bllb&b6a33gTHlg5XDM zE!EzId6r-N6Y3ka9PPwjcJ9sdt0qX(hmQ|AScWS{f&;^7+H)iAb;$#5dktvXWSvDm+ zV1;&bcHa6l#VMx`7{E+MusTOgLLSQK{07Xc_g2x|W!}Fv@_nV@0Amhpo*c~b1H+{T z^Fe&)iR`C&MWi&Zhs&!D>PjRK!qlpwOakXe;E04p2zuXo@sLR=7CS9i)y(3BmyKN-&o%gq1qH?<^;(SqUCcvQR2yN@R6(esF9} z+o6ceM&+yCtj;-i;cqiG)mO#ueFU4y^}4L6q-K_{)YWTYr`-H<_~RQAyz$|Gh7jF3 zZ?Z#jNw@4wa=l6~Z@+1I)E<5lKMKBv1w2#P*#E|bArZ^Kx)VphR6v03+e;lF>x@f9e{{#AcYYtw?hOS}sCuYUR|v2)3U`e8#;YI*sW6O(QA zQ3v!zdIgZR3&A|S8v|dx?It9 zoo&fbj}LNoF?m3Bp%D zl~5ng+M7ecrq#zgEBRY7SUOv6J(6-X*!Sr>kPnX9F!83(BbT+5Zsr_+F)nflX!n|< zSrFPZc2wCdJOR~c$kq! z_@^Pkq@A6^-!?y@$6=7b-4^1V(bBpHp{Q~wakCq=sU(9mAq=ZSm_RKmh50Vh%DOiT zvQQqpMsZ&Lb;@R;heC~eyLD4A6Qq>Vn)<{4!COQNSF%mQwHZ-%73ec%VeCFIq^BDr zy|G=Im){Y_U>$sycMygZ%-#Mf%fg;apFl~4?4Alywr-a@8zxD2r5K&^w-bq zNb}P@GI+EeQC%RYFeAXuGhyla6qjg_K!|&Ks^*p5<}0FZ?3+a=g`R%5_Na5Uag*^L zCi|qMmhj##u^1;wf0s$eaEfS!z7eIG&}AM9cMMU_60QnDM5Q&X3*T-7OmBt**74>u zyouWwK{$V?KBkl}+~J!IBpZ0Z3kkst67(1Ef7#{jDSt@RxX#NR1oUaqaGKfW9&Kk8 z6|)sr;f4fmZbZ2urLZq4;Ye^>?BY7yf;u*5gZxMiXz&Y3q_Q3-Ck{!)2DA0*E#}lkMJxW^Yg^Nk-$6E{w?s_ z-^wX&ZA3jgI<7lBCbby=@LrvApLaDzss+q`7}!%}bgnSB-)vCJD@8q6yNom=H6mBb zJt`CW>EnB=$u^IDk@n36h214=6ym1Ia)VGSYiCE&hRc0z8w1&V;@U&kI{ysLf%vCi zuSOie$-6cnKFi08>=pIu9%l`j^FF^m58_v&cm>8x1~9$_7}UU@FUCRK3v!dZuO>g}tgbpoV$Y}9x694Su-gLtz<*58-vq>~ zetm#gRyGd4XjN*OxDEYAlp$kK{0yN+!^!a>Y(>Y@Z0jDbk*cqu!U;1C;+H4|lT#Er zIALCr?dH7B`T{fHUT;tbgr}lB*B6bl2ct!7bB_;2STi9=hJGlxANidO%lLcfYgqw3o`%nHG+D^@un~IxgVIr9k3lZ|_{Qm!)luo6tO;D3lh~5l(VC;O!!txYJL^NTBw_F6^*e%iTg+P$idM|Q!}^VSv`x!Y(iW3 zFOlWsnYOk!=SIMQx+xS(W8?e1exOWh_mZBMs=H5l9&4ReT^?U3RLnfLpwQ>!gY+xQ zEv@ZUB-Jy|wpbx@s2xeI!}S+V{_qzX3QrzOwA5aR(zyH61r4%BSmRhqGpGLyAUlWL z+Ikjq#ZG=gzYCEAH2I)M2(^ty!k;cFOEWjH9gs4Om~Ao^<0c!v9nIIc8!k^A^V90Q z7;&oW5y>f(+HXfSG9}W!vh@EM&Eocv@uy{*d+7qsUGJjUhpvMr;QL9K1H1r=d>xcw z)3>!jF^{FUw@F49nm_FKxg!62S&@fopeLC~QsU~uAEM~S<4A;PJAGCQ7@K{#bXFqo z0SNHeXj(!0?&aF++9y=+ zbsuYMS`^xeuaVb0@l^eEoAxIAu(wi4E1PklR3RifHrzCbG~R_vLM6d)S2ZHB+Hsj^ zJSa_HeK+_W^=#7?u&u=auH~3cAYSo9`k8YCCp5ULzyS7e?)&TZ2y0Una4!_aF)F6i z-Qe77gU*1OxbhJwa`Ee;{HRDOqpYI}O*<{CqW6&URh}@%I$4rvTUP3;JbM>8vy{aS zcx#7+O0YvjyCG>0i>G|~u1h?A){@L6;A2;hSNrSh7XamF5Z|;_N8CM=^*l*(rF)vI zNZ+CE{fM;sjpi-ZNllAaUt>9mcW!V({kX*o#LH|zPdAKm@e;R-I@M|;srpCF~ z$oHn?D+IxSXla!@1Jnur&ycs98dpHb8P{BfVW|2_jN}8xYuNQ8kof{08axSWu(0bp z-b4vx#ZAV)AEcg@1Gj5C6P&g;m+H-~x+#U{TGHZI*NW`ptbNK~BAR!LZ{LL2Z|rI{ zhs!A(2p!u3GJI#=ULhZ=%V{x1ywXRG{;PG^K#%egcO#i}i^Jv@ zb+eF@iaR!YFshZ89%lL`9|qbK=3ZtHd>A6*!U5VN*ICT$sB{V(D9~tVU_5VZXEX0y z=uQryXhkHA#XI~sQi>ZtDfgilA)dJPbzHie{pR|ti$IYv&z;6<&QjFcp`$I9fR>+j zcCu&F(v4%JlV8k7PyD>OJ3F1y?)8{@*HdsGjYx_f=u2XAU?8H$sr8a6O+qC0u?9E0zF$GG!h@TZix*?3CQZU3 zK;=Awq%IfR4++x3>C#aSa%{fFeuPUmJG0dr8w=f$;aVbzaUZ-KYOK%+U`^l_uq%vS)mq zbf;~&r7{e=WaA6oW2ewy5_#pW9Q*}RN9U?!WU@pNaKLS@Xq?Aq6CwJ!2rQ;4pMxBkpg9++-F!)-8o zq2Xb|>_v)dbRqUr9~Ix&VmBiLHnC)$>YqN)_7NY2T>(L}>qQkXS=V@^+AOHXSM&iw0YcW1;I0$!VY@E?U#zv!(z(gioA#R5YBdu;{puaXgBj zNG62@s9T9?f8#i~#nyp;{Gc~|%t;!BCI2D}8Cq~0H?8)2m%XQPXL&U%*XgX1g$RR_ zxAglr5ZW2qKiuYU1FLHYfwtmez89QKxz8=uiGNTHIJex5*Vba6!`Nd%-*+og{%gr}N+4@vltShV z5T$eRMJ4YhGGtzBAeYNZpiHD0&O|gZL==?B?(J;hKW=}#WVG?^Y+jo_4n2Sj=tf=6 zU9*S3kkJ&=9WoPt_mk83v5~v}KL5BZcug$Ob%VM%R>cD*&^foSg4a(&O@X^-+i=q5Pa*}*jC;qe=O~1bbB__mJ_2$$DCV#@TI-nhIY7S zG;3hjFeFbXKT^?6_Pa|tPgp6YllK2v2>-Wf>7PD8p&he%bkeVuV&{r{s{{Jj+6`HR z%jF3;4Rirjt)B{*OlqT>1*4ai!HqZjB*F?di_8{woN&lHC>JSnIFfHsaj!WICDKP1 zG~G{*|6D6+kn=9W5n6f4@D>WH`+Sv0#uYX!uPgO>Movjs^FJaygfn zm&&-TN#h8C=ZgTjPf9ftFe-XdFlaas6N#kDu8&`GeO5j_-2_Ec;;PFhB4%aw%Fy=O3bTw#|~n1Y0QpSW*QY&0o9LG4qz9#Ww_`NV68v_UM*i(>Y@K%;bM zA`hnOMI$e%Nbe|;h&jV&EBMMU+&{@c1ZFVZ$q&d77>ko~9&Qx%2a_QvWNcV1-PfmD zL>Mo2tGGq>(8XF>lDcMQUBF(m_6NSNg=JYL+#KClD^g*@JlYVPGxjI`6Y+Q6$fxiGC{uCM&7&pCLKFdwUR=#c64+|#r>3diIJH13 zgKBF&(Y`hNp{5@G(u7NAOXC5vQfv3Le0}~XSjW5K+IkmNlQ%=y^RQK>O~DT2PPX3| zG0y^&dPHF_%n>BqM?nG?p#Jo{k1j{H0q`E0PY`5G65%>DF!oGvIIF5Hq4LT;_$(2J zMwu0sLBkJm0UKV+To|0+iTInMMcM8olZ%Dy=gFU5Sm7&S=1rI*WuWdRRlHpyqam`8|GqLW*3_+L;qpB6)E##6c@A<4$djr~P5q&3AoK9$Op&8gH7^ zBaH?y!Ab?5Vg;BYz@P@TD^BvQpidJp?nHA|zrGV& zHLJq&=2L9zm=O1AsZL%Z{TN0V8kdlz*CU%1ZE5MUYxmX}RP9a1ODgdJ^(tB$q;<~- z)o_#c--AlWCMG%s(hMETk(!-@WWuIRHC0AAj<}fcGu_|l-2bYJL4j!0+8Ybj4_7_euljz=>KC+@C#I=y8U6bE%NG&T5tPW0TMG|rC(YSC7qvm)FOykm!9g2jdC}Mcm|J0WM z*VFTX|AI^b2s;QOni>0T;E6%}W3+etsHX*zIZq;<*2H@Er#vr;IsS`5(g1)nyFN8D z;_c%PHug>aZT6AAz8@T!H=EB1KaHOpx9aH2+j44dIxu8d-@T_ojX0(J#rTyfLm+)8 zdyM1l+&Hx}(`#m8gm=#@hmjlVpYkjqXR~nEw%7z*^!rn0SlO|FzN@(@Lx82UVv zjhF1;pvC!bJP|_+Yh`9?Pn0*UoK9THj>{Pknhu5O0dieVbe15}bIcEp7tGj(q)Kt( z4IO;6bgOQ#;P!O$3ggjtDyrRsrsm;dfdGHC{(AyPZQ|7PL}oFjw(B2Z5gpZI38>zg zwJ`ykFcpG_{K(Uw%r{{AJYsHNry1nK9gZ+B_~nMDp76Xf`O5=~IZBrYv;D!jzO9sx zxa3S*WDn1Lj@yt~bTw)L=}EN*n>JVVc)5a2e1vIpg4kC|QizK45sW_+le78g#C;uq zO_o)dxjR=Q6Xl++%)JDa@)6|aIv8qTu#BV)#o=R(P1_dqmccs|-lSW7z37j?&+c^% z>ICgS;#8pH^{X^edofDwzG?F9xpZp-^3uF@|Sxub=0L z7ctKZIx#@rLnVll@Oi5KfJPo{uiYpD*>4ZvcUGqQ5*R3DcG-6!9|6VqC>Hh1EqZqPmb+8KC(I3&O0s8D%VIc-b*-*MfcSstWOy_ER@GuBWsHh6sk@ zRMlfz3Vd%PV={33H*vIo2o8#w|J$+u@7Cb|KZNNYx tuple[str, str]: return name, arguments -CAT_IMAGE_URL = "https://upload.wikimedia.org/wikipedia/commons/3/3a/Cat03.jpg" +_FIXTURES_DIR: Final = Path(__file__).parent / "fixtures" +CAT_IMAGE: Final = _FIXTURES_DIR / "cat.jpg" OPENAI_VISION_BACKEND = "openai/gpt-4o" -# OpenAI caches a shared prompt prefix once it exceeds ~1024 tokens; this is well -# past that, so a repeat call reports cached prompt tokens. + +def _cat_image_data_url() -> str: + """The vision image as a data URL, read from a fixture we own. + + An https URL would make every vision run depend on a third-party host staying + up and unthrottled, and a 429 from that host reads as a gateway failure. It also + changes what is under test per provider: litellm downloads the image itself for + bedrock, while openai is handed the link and fetches it from its own servers. A + data URL removes the host and puts both providers on the same bytes.""" + return "data:image/jpeg;base64," + base64.b64encode(CAT_IMAGE.read_bytes()).decode() + + def _vision_messages() -> list[ChatMessage]: return [ ChatMessage( role="user", content=[ TextContentPart(text="What animal is in this image? Answer in one word."), - ImageContentPart(image_url=ImageUrl(url=CAT_IMAGE_URL)), + ImageContentPart(image_url=ImageUrl(url=_cat_image_data_url())), ], ) ] From 40ff01b987932858e012262a2b2801f5ed16dd6f Mon Sep 17 00:00:00 2001 From: tin-berri Date: Thu, 27 Aug 2026 14:32:01 -0700 Subject: [PATCH 42/44] feat(mcp): let a resolved OAuth token target a custom upstream header (#38456) An MCP server behind an API gateway needs two credentials on one request: the gateway's own token on a private header, and a separate bearer on Authorization for the server behind it. Every arm that minted or held a token hardcoded Authorization, and the conflict rule then dropped the operator's static Authorization to make room, so the second credential never arrived. ApiKeyConfig already modelled this as header_name plus value_prefix behind a header() method. Extend that carrier to the four minted-token configs, have each resolver arm ask its config which header to use instead of naming one, and drop only the header the resolved credential is about to occupy. Operators set it per server via upstream_token_header, plumbed through config.yaml, the credentials blob, the management API and the admin form, on the M2M, token-exchange, authorization-code and ID-JAG arms. It is non-secret so it stays plaintext and round-trips on admin reads. Unset keeps today's behaviour. Moving a credential off Authorization means it stops inheriting what Authorization gets for free, so the slot now carries those protections itself. httpx drops Authorization when a redirect crosses origin and keeps every other header, so a custom slot is dropped by the client on the same condition, mirroring httpx's own scheme/host/port rule with an agreement test that fails if the two ever diverge. The v1 path also mirrors the v2 conflict rule, so an injected header cannot shadow the credential the gateway resolved for that slot. Which header a credential occupies, and what counts as being that header, was answered independently in nine places by four hand-rolled comparisons. same_header, has_header and without_header in litellm/types/mcp.py are now the one owner, shared by both MCP stacks, and the client derives its slot once instead of three times. The header name reaches egress verbatim, so the RFC 7230 grammar lives in one place and is checked where servers are built: a bad value fails the config load and the management API returns 400, rather than raising while a spec is built and emptying the aggregate tool list for every other server. A blank means unset, matching what the endpoint already accepts. --- litellm/experimental_mcp_client/client.py | 41 +++- .../mcp_server/mcp_server_manager.py | 46 ++--- .../mcp_server/oauth2_token_cache.py | 18 ++ .../mcp_server/openapi_to_mcp_generator.py | 40 +++- .../outbound_credentials/__init__.py | 6 + .../outbound_credentials/adapter.py | 17 +- .../client_credentials.py | 17 +- .../outbound_credentials/resolver.py | 19 +- .../mcp_server/outbound_credentials/types.py | 74 +++++-- .../proxy/_experimental/mcp_server/server.py | 4 +- .../mcp_management_endpoints.py | 19 ++ litellm/types/mcp.py | 101 +++++++++- .../types/mcp_server/mcp_server_manager.py | 19 +- .../test_mcp_client.py | 186 ++++++++++++++++++ .../outbound_credentials/test_adapter.py | 90 +++++++++ .../test_client_credentials.py | 65 +++++- .../outbound_credentials/test_resolver.py | 67 +++++++ .../outbound_credentials/test_types.py | 62 ++++++ .../mcp_server/test_mcp_server_manager.py | 138 ++++++++++++- .../mcp_server/test_oauth2_token_cache.py | 48 +++++ .../mcp_server/test_openapi_tool_auth.py | 118 +++++++++++ .../test_mcp_management_endpoints.py | 20 ++ tests/test_litellm/types/test_mcp.py | 87 ++++++++ .../_components/IdJagFormFields.tsx | 2 + .../_components/OAuthFormFields.test.tsx | 38 ++++ .../_components/OAuthFormFields.tsx | 3 + .../_components/TokenExchangeFormFields.tsx | 2 + .../_components/UpstreamTokenHeaderField.tsx | 31 +++ .../editServerPayload.differential.cases.ts | 19 +- .../_components/mountedServerFields.test.ts | 32 ++- .../_components/mountedServerFields.ts | 5 +- .../src/components/mcp_tools/types.tsx | 2 +- 32 files changed, 1349 insertions(+), 87 deletions(-) create mode 100644 tests/test_litellm/types/test_mcp.py create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/UpstreamTokenHeaderField.tsx diff --git a/litellm/experimental_mcp_client/client.py b/litellm/experimental_mcp_client/client.py index be9d2b88e99..f0a1bff8fdc 100644 --- a/litellm/experimental_mcp_client/client.py +++ b/litellm/experimental_mcp_client/client.py @@ -56,6 +56,9 @@ from litellm.types.mcp import ( MCPStdioConfig, MCPTransport, MCPTransportType, + credential_redirect_hook, + has_header, + without_header, ) @@ -273,6 +276,7 @@ class MCPClient: transport_type: MCPTransportType = MCPTransport.http, auth_type: MCPAuthType = None, auth_value: str | dict[str, str] | None = None, + auth_header_name: str | None = None, timeout: float | None = None, stdio_config: MCPStdioConfig | None = None, extra_headers: dict[str, str] | None = None, @@ -288,6 +292,11 @@ class MCPClient: self.auth_type: MCPAuthType = auth_type self.timeout: float = timeout if timeout is not None else MCP_CLIENT_TIMEOUT self._mcp_auth_value: str | dict[str, str] | None = None + # The one place this client decides which header its credential occupies: the operator's + # configured slot on the v1 path, or the slot the v2 resolver's auth object already owns. + # Every consumer reads this rather than re-deriving it, since each re-derivation so far + # picked up a different bug. + self._credential_slot: str | None = auth_header_name or getattr(resolved_auth, "header_name", None) self.stdio_config: MCPStdioConfig | None = stdio_config self.extra_headers: dict[str, str] | None = extra_headers self.ssl_verify: VerifyTypes | None = ssl_verify @@ -501,26 +510,33 @@ class MCPClient: else: self._mcp_auth_value = mcp_auth_value + def _header_slot(self, default: str) -> str: + return self._credential_slot or default + def _get_auth_headers(self) -> dict: """Generate authentication headers based on auth type.""" headers: Final = {} if self._mcp_auth_value: if isinstance(self._mcp_auth_value, str): if self.auth_type == MCPAuth.bearer_token: - headers["Authorization"] = f"Bearer {strip_auth_scheme(self._mcp_auth_value, 'Bearer')}" + static_bearer: Final = strip_auth_scheme(self._mcp_auth_value, "Bearer") + headers[self._header_slot("Authorization")] = f"Bearer {static_bearer}" elif self.auth_type == MCPAuth.basic: - headers["Authorization"] = f"Basic {self._mcp_auth_value}" + headers[self._header_slot("Authorization")] = f"Basic {self._mcp_auth_value}" elif self.auth_type == MCPAuth.api_key: - headers["X-API-Key"] = self._mcp_auth_value + headers[self._header_slot("X-API-Key")] = self._mcp_auth_value elif self.auth_type == MCPAuth.authorization: # This auth type means the caller owns the whole header value. - headers["Authorization"] = self._mcp_auth_value + headers[self._header_slot("Authorization")] = self._mcp_auth_value elif self.auth_type == MCPAuth.oauth2: - headers["Authorization"] = f"Bearer {strip_auth_scheme(self._mcp_auth_value, 'Bearer')}" + oauth2_bearer: Final = strip_auth_scheme(self._mcp_auth_value, "Bearer") + headers[self._header_slot("Authorization")] = f"Bearer {oauth2_bearer}" elif self.auth_type == MCPAuth.token: - headers["Authorization"] = f"token {strip_auth_scheme(self._mcp_auth_value, 'token')}" + scheme_token: Final = strip_auth_scheme(self._mcp_auth_value, "token") + headers[self._header_slot("Authorization")] = f"token {scheme_token}" elif self.auth_type == MCPAuth.oauth2_token_exchange: - headers["Authorization"] = f"Bearer {strip_auth_scheme(self._mcp_auth_value, 'Bearer')}" + exchanged_bearer: Final = strip_auth_scheme(self._mcp_auth_value, "Bearer") + headers[self._header_slot("Authorization")] = f"Bearer {exchanged_bearer}" elif isinstance(self._mcp_auth_value, dict): headers.update(self._mcp_auth_value) # Note: aws_sigv4 auth is not handled here — SigV4 requires per-request @@ -528,7 +544,14 @@ class MCPClient: # of static headers. See MCPSigV4Auth and _create_httpx_client_factory(). # update the headers with the extra headers if self.extra_headers: - headers.update(self.extra_headers) + # Mirrors _resolve_v2_auth: when the operator named a slot for the credential the + # gateway resolved, no injected header may shadow it, case-insensitively, since HTTP + # header names are. Without a configured slot the old precedence stands unchanged. + slot: Final = self._credential_slot + injected: Final = ( + without_header(self.extra_headers, slot) if slot and has_header(headers, slot) else self.extra_headers + ) + headers.update(injected or {}) return _strip_header_whitespace(headers) def _create_httpx_client_factory(self) -> Callable[..., httpx.AsyncClient]: @@ -556,12 +579,14 @@ class MCPClient: # SigV4 aws_auth. Both are None for the common case — no behavior change. fallback_auth: Final = self._resolved_auth if self._resolved_auth is not None else self._aws_auth effective_auth: Final = auth if auth is not None else fallback_auth + guard: Final = credential_redirect_hook(self.server_url, self._credential_slot) return httpx.AsyncClient( headers=headers, timeout=timeout, auth=effective_auth, verify=ssl_config, follow_redirects=True, + event_hooks={"request": [guard]} if guard else {}, ) return factory diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py index 6a1b6851d3e..2330120adad 100644 --- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py +++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py @@ -34,6 +34,7 @@ from mcp.types import ( ) from mcp.types import Tool as MCPTool from pydantic import AnyUrl, BaseModel +from typing_extensions import ReadOnly import litellm from litellm._logging import verbose_logger @@ -72,6 +73,7 @@ from litellm.proxy._experimental.mcp_server.oauth2_token_cache import ( MCPPerUserTokenCache, mcp_per_user_token_cache, resolve_mcp_auth, + resolved_token_header, ) from litellm.proxy._experimental.mcp_server.oauth_utils import ( _redact_mcp_resource_url, @@ -99,6 +101,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.token_exchange_ build_token_exchanger, ) from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( + DEFAULT_CREDENTIAL_HEADER, AuthorizationCodeConfig, ClientCredentialsConfig, CredError, @@ -153,6 +156,8 @@ from litellm.types.mcp import ( MCPAuth, MCPStdioConfig, MCPTokenEndpointAuthMethod, + has_header, + without_header, ) from litellm.types.mcp_server.mcp_server_manager import ( MCPInfo, @@ -349,6 +354,7 @@ class MCPServerConfig(TypedDict, total=False): audience: str subject_token_type: str upstream_resource: str + upstream_token_header: ReadOnly[str] id_jag_resource_token_endpoint: str id_jag_resource: str client_private_key: str @@ -828,18 +834,6 @@ def _should_strip_caller_authorization( ) -def _without_authorization( - headers: dict[str, str] | None, -) -> dict[str, str] | None: - """A copy of ``headers`` with any ``Authorization`` key removed (case-insensitive), or - None if nothing remains. Drops only the credential, keeping other forwarded headers. - """ - if not headers: - return None - filtered: Final = {k: v for k, v in headers.items() if k.lower() != "authorization"} - return filtered or None - - def _format_byok_openapi_auth_header(mcp_server: MCPServer, mcp_auth_header: str) -> str: """Format a raw BYOK credential for OpenAPI tool ``Authorization`` injection. @@ -914,7 +908,9 @@ def _resolve_openapi_tool_auth( if isinstance(per_server, dict): authorization: Final = next((v for k, v in per_server.items() if k.lower() == "authorization"), None) - merged: Final = merge_mcp_headers(extra_headers=forwarded, static_headers=_without_authorization(per_server)) + merged: Final = merge_mcp_headers( + extra_headers=forwarded, static_headers=without_header(per_server, DEFAULT_CREDENTIAL_HEADER) + ) if authorization is None: byok: Final = _format_byok_openapi_auth_header(mcp_server, mcp_auth_header) if mcp_auth_header else None return byok, merged, mcp_auth_header @@ -981,7 +977,7 @@ def _client_forwarded_authorization_headers( raw_headers=raw_headers, user_api_key_auth=user_api_key_auth, ): - return _without_authorization(extra_headers) + return without_header(extra_headers, DEFAULT_CREDENTIAL_HEADER) return extra_headers @@ -994,7 +990,7 @@ def _take_forwarded_authorization( if not headers: return None, headers value: Final = next((v for k, v in headers.items() if k.lower() == "authorization"), None) - return value, _without_authorization(headers) + return value, without_header(headers, DEFAULT_CREDENTIAL_HEADER) def _passthrough_token_from_mcp_auth_header( @@ -2166,6 +2162,7 @@ class MCPServerManager: DEFAULT_SUBJECT_TOKEN_TYPE, ), upstream_resource=server_config.get("upstream_resource", None), + upstream_token_header=server_config.get("upstream_token_header", None), # ID-JAG fields id_jag_resource_token_endpoint=server_config.get("id_jag_resource_token_endpoint", None), id_jag_resource=server_config.get("id_jag_resource", None), @@ -2698,6 +2695,7 @@ class MCPServerManager: or (credentials_dict.get("subject_token_type") if credentials_dict else None) or DEFAULT_SUBJECT_TOKEN_TYPE, upstream_resource=(credentials_dict.get("upstream_resource") if credentials_dict else None), + upstream_token_header=(credentials_dict.get("upstream_token_header") if credentials_dict else None), # ID-JAG fields — read from credentials JSON blob id_jag_resource_token_endpoint=( credentials_dict.get("id_jag_resource_token_endpoint") if credentials_dict else None @@ -3525,10 +3523,9 @@ class MCPServerManager: case Ok(auth): # NoOpAuth has no header_name and so never conflicts. header_name: Final[str | None] = getattr(auth, "header_name", None) - conflicts: Final = bool( - header_name and extra_headers and any(key.lower() == header_name.lower() for key in extra_headers) - ) - if not conflicts: + if header_name is None or not extra_headers: + return auth, extra_headers + if not has_header(extra_headers, header_name): return auth, extra_headers if isinstance( spec.config, @@ -3540,9 +3537,10 @@ class MCPServerManager: # guardrail such as MCPJWTSigner, static_headers, or any other injected # Authorization must NOT shadow it (otherwise the upstream gets e.g. the # signer's JWT instead of the minted token and rejects it, and for M2M the - # one-shot 401 refetch is lost with it). Drop the conflicting header so the - # resolved token reaches upstream. - return auth, _without_authorization(extra_headers) + # one-shot 401 refetch is lost with it). Drop only the header the resolved + # credential is about to occupy, so a static credential the operator aimed at a + # DIFFERENT header still reaches upstream. + return auth, without_header(extra_headers, header_name) # Other modes: an Authorization already supplied via extra_headers (a forwarded caller # header or static_headers) is intentional and wins; v1 applies those last. return None, extra_headers @@ -3650,6 +3648,7 @@ class MCPServerManager: ): spec = None auth_value: Final = await resolve_mcp_auth(resolved_server, mcp_auth_header) if spec is None else None + auth_header_name: Final = resolved_token_header(resolved_server, mcp_auth_header) if spec is None else None # Create sampling and elicitation callbacks for this client sampling_cb = ( @@ -3758,6 +3757,7 @@ class MCPServerManager: transport_type=transport, auth_type=resolved_server.auth_type, auth_value=auth_value, + auth_header_name=auth_header_name, timeout=(resolved_server.timeout if resolved_server.timeout is not None else MCP_CLIENT_TIMEOUT), extra_headers=extra_headers, aws_auth=aws_auth, @@ -5306,7 +5306,7 @@ class MCPServerManager: raw_headers=raw_headers, user_api_key_auth=user_api_key_auth, ): - extra_headers = _without_authorization(extra_headers) + extra_headers = without_header(extra_headers, DEFAULT_CREDENTIAL_HEADER) elif mcp_server.is_client_forwarded_token: extra_headers = _client_forwarded_authorization_headers( mcp_server=mcp_server, diff --git a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py index b3f1da51074..a4ef970b87a 100644 --- a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py +++ b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py @@ -7,6 +7,7 @@ with ``client_id``, ``client_secret``, and ``token_url``. import asyncio import hashlib +from collections.abc import Mapping from typing import TYPE_CHECKING, Final import httpx @@ -313,9 +314,26 @@ async def resolve_mcp_auth( 1. ``mcp_auth_header`` — per-request/per-user override 2. OAuth2 client_credentials token — auto-fetched and cached 3. ``server.authentication_token`` — static token from config/DB + + ``resolved_token_header`` answers, for the same two inputs, which header the value belongs in. """ if mcp_auth_header: return mcp_auth_header if server.has_client_credentials: return await mcp_oauth2_token_cache.async_get_token(server) return server.authentication_token + + +def resolved_token_header( + server: "MCPServer", + mcp_auth_header: str | Mapping[str, str] | None = None, +) -> str | None: + """Which upstream header the value ``resolve_mcp_auth`` just returned belongs in. + + ``None`` means keep the auth_type default. A caller-supplied ``mcp_auth_header`` is the caller's + own credential aimed at the slot the upstream normally uses, so it never moves; only the values + the gateway resolved from its own config (the minted M2M token, the static token) follow + ``upstream_token_header``. Same inputs and same branch order as ``resolve_mcp_auth``, so the two + cannot disagree about which case they are in. + """ + return None if mcp_auth_header else server.upstream_token_header diff --git a/litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py b/litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py index 083a98cdd36..16f58ef5b76 100644 --- a/litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py +++ b/litellm/proxy/_experimental/mcp_server/openapi_to_mcp_generator.py @@ -47,12 +47,14 @@ def sanitize_openapi_tool_name(raw_name: str) -> str: from litellm._logging import verbose_logger from litellm.litellm_core_utils.url_utils import async_safe_get from litellm.llms.custom_httpx.http_handler import ( + AsyncHTTPHandler, get_async_httpx_client, httpxSpecialProvider, ) from litellm.proxy._experimental.mcp_server.tool_registry import ( global_mcp_tool_registry, ) +from litellm.types.mcp import credential_redirect_hook, custom_credential_slot class _OpenAPIJSONSchema(TypedDict, total=False): @@ -119,6 +121,10 @@ _request_resolved_auth_headers: Final[contextvars.ContextVar[dict[str, str] | No "_request_resolved_auth_headers", default=None ) +_request_upstream_url: Final[contextvars.ContextVar[str | None]] = contextvars.ContextVar( + "_request_upstream_url", default=None +) + def _sanitize_path_parameter_value(param_value: object, param_name: str) -> str: """Ensure path params cannot introduce directory traversal.""" @@ -349,6 +355,35 @@ def build_input_schema(operation: _OpenAPIOperation) -> dict[str, object]: } +async def _drop_credential_across_origin(request: httpx.Request) -> None: + """Apply this request's cross-origin credential guard, if it needs one. + + Reads the per-request context rather than closing over it so the hook is one stable object, which + keeps the guarded client cacheable. A closure would key a new entry per call, and the handler it + built would never be closed. + """ + guard: Final = credential_redirect_hook( + _request_upstream_url.get() or "", custom_credential_slot(_request_resolved_auth_headers.get()) + ) + if guard is not None: + await guard(request) + + +def _upstream_client() -> AsyncHTTPHandler: + """The HTTP client for one upstream call, guarded when a credential rides a custom slot. + + A resolved credential outside ``Authorization`` is not stripped across origins by the client + itself, so this arm installs the same hook the MCP client uses. Both variants come from the + shared cache, so a guarded call reuses its connection pool like any other. + """ + if custom_credential_slot(_request_resolved_auth_headers.get()) is None: + return get_async_httpx_client(llm_provider=httpxSpecialProvider.MCP) + return get_async_httpx_client( + llm_provider=httpxSpecialProvider.MCP, + params={"event_hooks": {"request": [_drop_credential_across_origin]}}, + ) + + def _merge_openapi_tool_request_headers( static_headers: dict[str, str], ) -> dict[str, str]: @@ -510,8 +545,9 @@ def create_tool_function( except (json.JSONDecodeError, TypeError): json_body = {"data": body_value} - client: Final = get_async_httpx_client(llm_provider=httpxSpecialProvider.MCP) + client: Final = _upstream_client() upstream: Final = server_label or f"{original_method.upper()} {path}" + url_token: Final = _request_upstream_url.set(url) try: if original_method == "get": @@ -529,6 +565,8 @@ def create_tool_function( except MaskedHTTPStatusError as e: _raise_for_upstream_failure(e.response, upstream, relays_upstream_auth) raise + finally: + _request_upstream_url.reset(url_token) _raise_for_upstream_failure(response, upstream, relays_upstream_auth) return response.text diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/__init__.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/__init__.py index a5dc75e3829..d61f8395677 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/__init__.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/__init__.py @@ -21,6 +21,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.result import ( Result, ) from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( + DEFAULT_CREDENTIAL_HEADER, Ambient, ApiKeyConfig, ApiKeySource, @@ -35,6 +36,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( ClientCredentialsConfig, ClientSecretAuth, CredError, + HeaderCarrier, IdJagConfig, NoneConfig, PassthroughConfig, @@ -45,9 +47,11 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( Subject, TokenExchangeConfig, parse_auth_spec_kind, + validate_header_name, ) __all__ = [ + "DEFAULT_CREDENTIAL_HEADER", "Ambient", "ApiKeyConfig", "ApiKeySource", @@ -63,6 +67,7 @@ __all__ = [ "ClientSecretAuth", "CredError", "Error", + "HeaderCarrier", "IdJagConfig", "NoOpAuth", "NoneConfig", @@ -78,4 +83,5 @@ __all__ = [ "TokenExchangeConfig", "UpstreamCredentialProvider", "parse_auth_spec_kind", + "validate_header_name", ] diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py index 98e239b1d1d..4458ac7f190 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py @@ -20,6 +20,7 @@ from typing_extensions import assert_never from litellm.proxy._experimental.mcp_server.oauth_utils import resolve_upstream_resource from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( + DEFAULT_CREDENTIAL_HEADER, ApiKeyConfig, AuthorizationCodeConfig, ClientAuth, @@ -45,6 +46,15 @@ _TOKEN_EXCHANGE_SUBJECT_TOKEN_DEFAULT: Final = "urn:ietf:params:oauth:token-type _ID_JAG_SUBJECT_TOKEN_DEFAULT: Final = "urn:ietf:params:oauth:token-type:id_token" +def token_header(server: MCPServer) -> str: + """The upstream header this server's resolved credential occupies. + + One owner for every arm, so no spec builder spells the default itself and a server can never + hand two arms different answers. + """ + return server.upstream_token_header or DEFAULT_CREDENTIAL_HEADER + + def to_subject(user_api_key_auth: UserAPIKeyAuth | None, subject_token: str | None) -> Subject: """Map v1's authenticated principal onto the resolver's Subject. @@ -122,7 +132,7 @@ def _oauth2_spec(server: MCPServer, resource: str) -> ServerSpec | None: return ServerSpec( server_id=server.server_id, resource=resource, - config=AuthorizationCodeConfig(), + config=AuthorizationCodeConfig(header_name=token_header(server)), ) return None @@ -140,6 +150,7 @@ def _client_credentials_spec(server: MCPServer, resource: str) -> ServerSpec: server_id=server.server_id, resource=resource, config=ClientCredentialsConfig( + header_name=token_header(server), client_id=server.client_id, client_secret=SecretStr(server.client_secret) if server.client_secret else None, token_url=server.effective_token_url, @@ -173,6 +184,7 @@ def _token_exchange_spec(server: MCPServer, resource: str) -> ServerSpec | None: server_id=server.server_id, resource=resource, config=TokenExchangeConfig( + header_name=token_header(server), profile=profile, subject_token_type=server.subject_token_type or DEFAULT_SUBJECT_TOKEN_TYPE, token_exchange_endpoint=endpoint, @@ -206,7 +218,7 @@ def _shared_key_spec( server_id=server.server_id, resource=resource, config=ApiKeyConfig( - header_name=header_name, + header_name=server.upstream_token_header or header_name, value_prefix=value_prefix, key_source=SharedKey(value=SecretStr(value)), ), @@ -231,6 +243,7 @@ def _id_jag_spec(server: MCPServer, resource: str) -> ServerSpec | None: server_id=server.server_id, resource=resource, config=IdJagConfig( + header_name=token_header(server), org_token_endpoint=org_token_endpoint, resource_token_endpoint=resource_token_endpoint, client_id=client_id, diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/client_credentials.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/client_credentials.py index d0053fbe0a8..da00abfe604 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/client_credentials.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/client_credentials.py @@ -50,6 +50,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.result import ( from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( ClientCredentialsConfig, CredError, + HeaderCarrier, ) @@ -328,14 +329,21 @@ class ClientCredentialsBearerAuth(httpx.Auth): refetch fails, or the retried request 401s again, the upstream's response stands. """ - def __init__(self, access_token: str, refetch: Callable[[str], Awaitable[str | None]]) -> None: - self.header_name = "Authorization" + def __init__( + self, + access_token: str, + refetch: Callable[[str], Awaitable[str | None]], + carrier: HeaderCarrier, + ) -> None: + self._carrier = carrier + self.header_name = carrier.header_name self._access_token = SecretStr(access_token) self._refetch = refetch async def async_auth_flow(self, request: httpx.Request) -> AsyncGenerator[httpx.Request, httpx.Response]: token: Final = self._access_token.get_secret_value() - request.headers[self.header_name] = f"Bearer {token}" + name, value = self._carrier.header(token) + request.headers[name] = value response: Final = yield request if response.status_code != 401: return @@ -343,7 +351,8 @@ class ClientCredentialsBearerAuth(httpx.Auth): if fresh is None: return self._access_token = SecretStr(fresh) - request.headers[self.header_name] = f"Bearer {fresh}" + fresh_name, fresh_value = self._carrier.header(fresh) + request.headers[fresh_name] = fresh_value yield request def sync_auth_flow(self, request: httpx.Request) -> Generator[httpx.Request, httpx.Response, None]: diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py index 94c59962b70..3af7b51f432 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py @@ -145,8 +145,8 @@ class UpstreamCredentialProvider: return await self._token_exchange(subject, server, config) case IdJagConfig() as config: return await self._id_jag(subject, server, config) - case AuthorizationCodeConfig(): - return await self._authorization_code(subject, server) + case AuthorizationCodeConfig() as config: + return await self._authorization_code(subject, server, config) case AwsSigV4Config(): return _not_implemented(AuthSpecKind.aws_sigv4) assert_never(server.config) @@ -284,15 +284,19 @@ class UpstreamCredentialProvider: match await self._exchanged_tokens.get_or_compute(slot, _exchange, fingerprint=fingerprint): case Ok(access_token): - return Ok(StaticHeaderAuth(f"Bearer {access_token}")) + header_name, header_value = config.header(access_token) + return Ok(StaticHeaderAuth(header_value, header_name=header_name)) case Error(err): return Error(err) - async def _authorization_code(self, subject: Subject, server: ServerSpec) -> Result[StaticHeaderAuth, CredError]: + async def _authorization_code( + self, subject: Subject, server: ServerSpec, config: AuthorizationCodeConfig + ) -> Result[StaticHeaderAuth, CredError]: token: Final = await self._authz_token(subject, server) if token is None: return Error(CredError.of_unauthorized("Authorization required: complete the OAuth flow for this server.")) - return Ok(StaticHeaderAuth(f"Bearer {token.access_token}", header_name="Authorization")) + header_name, header_value = config.header(token.access_token) + return Ok(StaticHeaderAuth(header_value, header_name=header_name)) async def _client_credentials( self, server_id: str, config: ClientCredentialsConfig @@ -307,7 +311,7 @@ class UpstreamCredentialProvider: match await self._client_credentials_source.get(server_id, config): case Ok(token): refetch: Final = partial(self._client_credentials_source.refetch, server_id, config) - return Ok(ClientCredentialsBearerAuth(token.access_token, refetch)) + return Ok(ClientCredentialsBearerAuth(token.access_token, refetch, config)) case Error(err): return Error(err) @@ -332,7 +336,8 @@ class UpstreamCredentialProvider: inbound.get_secret_value(), server, config, tenant_id=subject.tenant_id ): case Ok(token): - return Ok(StaticHeaderAuth(f"Bearer {token.access_token}", header_name="Authorization")) + header_name, header_value = config.header(token.access_token) + return Ok(StaticHeaderAuth(header_value, header_name=header_name)) case Error(err): return Error(err) diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py index ce9948f0448..67aad3e443e 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py @@ -31,7 +31,7 @@ from enum import Enum from typing import Annotated, Final, Literal from expression import case, tag, tagged_union -from pydantic import BaseModel, ConfigDict, Field, SecretStr +from pydantic import BaseModel, ConfigDict, Field, SecretStr, field_validator from typing_extensions import assert_never from litellm.proxy._experimental.mcp_server.outbound_credentials.result import ( @@ -39,7 +39,11 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.result import ( Ok, Result, ) -from litellm.types.mcp import DEFAULT_SUBJECT_TOKEN_TYPE +from litellm.types.mcp import ( + DEFAULT_CREDENTIAL_HEADER, + DEFAULT_SUBJECT_TOKEN_TYPE, + normalize_upstream_header_name, +) class AuthSpecKind(str, Enum): @@ -161,7 +165,52 @@ class CredError: assert_never(self.tag) -class AuthorizationCodeConfig(BaseModel): +def validate_header_name(raw: str) -> Result[str, CredError]: + """``normalize_upstream_header_name`` with this package's error-as-value policy. + + The grammar itself lives in ``litellm.types.mcp`` so the v1 model, the management endpoint and + this vocabulary all judge a header name the same way while each keeps its own failure shape. + """ + normalized: Final = normalize_upstream_header_name(raw) + if normalized is None: + return Error(CredError.of_misconfigured(f"invalid upstream header name: {raw!r}")) + return Ok(normalized) + + +class HeaderCarrier(BaseModel): + """Where a resolved credential is written upstream, and how its value is formatted. + + ``Authorization: Bearer`` is only OAuth's *default* conveyance (RFC 6750 section 2.1), not its + only one: an ESB or API gateway commonly terminates its own credential in a private header while + a second credential passes through to the origin, so a credential has to be able to say which + slot it owns. Modeled like OpenAPI's apiKey scheme, so any upstream convention is expressible + (Authorization + Bearer, a raw value on X-API-Key, Ocp-Apim-Subscription-Key, esb-oauth, ...). + + Every config whose credential the gateway mints or holds inherits this, so no resolver arm names + a header itself and the conflict rule in ``_resolve_v2_auth`` can always ask the auth object + which slot it is about to occupy. ``passthrough`` deliberately does not: it forwards the + caller's own credential into the slot the caller used, and mints nothing to place. + """ + + model_config = ConfigDict(frozen=True) + header_name: str = DEFAULT_CREDENTIAL_HEADER + value_prefix: str = "Bearer" + + @field_validator("header_name") + @classmethod + def _check_header_name(cls, value: str) -> str: + match validate_header_name(value): + case Ok(name): + return name + case Error(err): + raise ValueError(err.summary) + + def header(self, value: str) -> tuple[str, str]: + formatted: Final = f"{self.value_prefix} {value}" if self.value_prefix else value + return self.header_name, formatted + + +class AuthorizationCodeConfig(HeaderCarrier): """Per-user 3LO; the gateway is the OAuth client and stores the user's token. Endpoints are discovered (RFC 9728 -> RFC 8414) and the client is registered via DCR @@ -179,7 +228,7 @@ class AuthorizationCodeConfig(BaseModel): token_url: str | None = None -class ClientCredentialsConfig(BaseModel): +class ClientCredentialsConfig(HeaderCarrier): """M2M service account; one upstream identity for every user. Fields are optional so the config can be built incomplete: a value may be supplied at @@ -203,7 +252,7 @@ class ClientCredentialsConfig(BaseModel): token_endpoint_auth_method: Literal["client_secret_post", "client_secret_basic"] | None = None -class TokenExchangeConfig(BaseModel): +class TokenExchangeConfig(HeaderCarrier): """OBO: swap the caller's live inbound token for a token bound to the upstream's audience. The gateway authenticates to the exchange endpoint as an OAuth client (`client_id`/`client_secret`); the inbound token is sent only to that endpoint, never to the upstream. @@ -255,7 +304,7 @@ class ClientSecretAuth(BaseModel): ClientAuth = Annotated[PrivateKeyJwtAuth | ClientSecretAuth, Field(discriminator="source")] -class IdJagConfig(BaseModel): +class IdJagConfig(HeaderCarrier): """draft-ietf-oauth-identity-assertion-authz-grant (Okta "AI agent token exchange"). Two legs: leg 1 is an RFC 8693 token exchange at the IdP org AS (`org_token_endpoint`) that @@ -297,23 +346,16 @@ class Byok(BaseModel): ApiKeySource = Annotated[SharedKey | Byok, Field(discriminator="source")] -class ApiKeyConfig(BaseModel): +class ApiKeyConfig(HeaderCarrier): """A fixed credential injected as a header. The value is shared (in config) or seeded - per-user (pulled from the store); `header_name` and `value_prefix` say where and how it is - written, modeled like OpenAPI's apiKey scheme so any upstream convention is expressible - (Authorization + Bearer, a raw value on X-API-Key, Ocp-Apim-Subscription-Key, etc.). + per-user (pulled from the store); the inherited `header_name` and `value_prefix` say where + and how it is written. """ model_config = ConfigDict(frozen=True) kind: Literal[AuthSpecKind.api_key] = AuthSpecKind.api_key - header_name: str = "Authorization" - value_prefix: str = "Bearer" key_source: ApiKeySource - def header(self, value: str) -> tuple[str, str]: - formatted: Final = f"{self.value_prefix} {value}" if self.value_prefix else value - return self.header_name, formatted - class PassthroughConfig(BaseModel): """Client-driven upstream OAuth; the gateway forwards the client's upstream token.""" diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py index 57e59dab2d1..c6b2ac489bb 100644 --- a/litellm/proxy/_experimental/mcp_server/server.py +++ b/litellm/proxy/_experimental/mcp_server/server.py @@ -433,7 +433,6 @@ if MCP_AVAILABLE: _client_forwarded_authorization_headers, _resolve_openapi_tool_auth, _should_strip_caller_authorization, - _without_authorization, global_mcp_server_manager, ) from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import ( @@ -452,6 +451,7 @@ if MCP_AVAILABLE: split_server_prefix_from_name, strip_known_server_prefix, ) + from litellm.types.mcp import DEFAULT_CREDENTIAL_HEADER, without_header ###################################################### ############ MCP Tools List REST API Response Object # @@ -1733,7 +1733,7 @@ if MCP_AVAILABLE: raw_headers=raw_headers, user_api_key_auth=user_api_key_auth, ): - extra_headers = _without_authorization(extra_headers) + extra_headers = without_header(extra_headers, DEFAULT_CREDENTIAL_HEADER) elif is_client_forwarded_mode: if not withhold_forwarded_authorization: extra_headers = _client_forwarded_authorization_headers( diff --git a/litellm/proxy/management_endpoints/mcp_management_endpoints.py b/litellm/proxy/management_endpoints/mcp_management_endpoints.py index 54a591a5e1a..556a30d0b29 100644 --- a/litellm/proxy/management_endpoints/mcp_management_endpoints.py +++ b/litellm/proxy/management_endpoints/mcp_management_endpoints.py @@ -204,6 +204,7 @@ if MCP_AVAILABLE: MCP_ADMIN_CONFIG_CREDENTIAL_KEYS, MCPAuth, MCPCredentials, + normalize_upstream_header_name, ) from litellm.types.mcp_server.mcp_server_manager import MCPServer @@ -239,9 +240,26 @@ if MCP_AVAILABLE: detail={"error": error_messages_text}, ) + def _validate_upstream_token_header(payload: McpServerPayloadLike) -> None: + credentials: Final = getattr(payload, "credentials", None) + raw: Final = credentials.get("upstream_token_header") if isinstance(credentials, dict) else None + if not isinstance(raw, str) or raw == "": + return + if normalize_upstream_header_name(raw) is None: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail={ + "error": ( + f"Invalid upstream_token_header {raw!r}: must be a valid HTTP header name " + "(RFC 7230 token, e.g. 'esb-oauth')" + ) + }, + ) + def validate_and_normalize_mcp_server_payload(payload: McpServerPayloadLike) -> None: _base_validate_and_normalize_mcp_server_payload(payload) _validate_mcp_server_name_fields(payload) + _validate_upstream_token_header(payload) def stamp_omitted_oauth2_flow(payload: NewMCPServerRequest) -> None: """Fallback only: fill in oauth2_flow when an oauth2 create omits it. @@ -739,6 +757,7 @@ if MCP_AVAILABLE: ("aws_region_name", "aws_region_name"), ("aws_service_name", "aws_service_name"), ("upstream_resource", "upstream_resource"), + ("upstream_token_header", "upstream_token_header"), ) def _has_non_admin_config_credentials(credentials: "MCPCredentials | None") -> bool: diff --git a/litellm/types/mcp.py b/litellm/types/mcp.py index 57437ea7e54..1b8baf2da09 100644 --- a/litellm/types/mcp.py +++ b/litellm/types/mcp.py @@ -1,6 +1,11 @@ import enum +import re +from collections.abc import Awaitable, Callable, Mapping +from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal +from urllib.parse import urlsplit +import httpx from pydantic import BaseModel from typing_extensions import TypedDict @@ -181,6 +186,15 @@ class MCPCredentials(TypedDict, total=False): ``audience``, which is the RFC 8693 token-exchange parameter. """ + upstream_token_header: str | None # writable-ok: pydantic warns it cannot honour ReadOnly here + """ + Which upstream header carries the credential LiteLLM resolves for this server. Omitted when + unset, which keeps RFC 6750's default of ``Authorization``. Set it when the upstream expects the + gateway's token somewhere else (an ESB terminating its own credential on e.g. ``esb-oauth``), so + a separate operator-configured ``Authorization`` reaches the origin untouched. Non-secret, so it + is stored in plaintext and returned on admin reads. + """ + client_private_key: str | None """ PEM private key used to sign the private-key-JWT client_assertion (RFC 7523) @@ -223,7 +237,92 @@ class MCPCredentials(TypedDict, total=False): """ -MCP_ADMIN_CONFIG_CREDENTIAL_KEYS: Final[tuple[str, ...]] = ("upstream_resource",) +DEFAULT_CREDENTIAL_HEADER: Final = "Authorization" + +_HEADER_NAME_TOKEN: Final = re.compile(r"^[!#$%&'*+\-.^_`|~0-9A-Za-z]+$") + + +def normalize_upstream_header_name(raw: str) -> str | None: + """The trimmed header name if it is a usable RFC 7230 ``token``, else None. + + One owner for the grammar; each caller picks its own failure shape (a config-load raise, an + API 400, a typed CredError). An operator-supplied name reaches egress verbatim, so a value + carrying CR/LF, spaces or separators must never get that far. + """ + stripped: Final = raw.strip() + return stripped if stripped and _HEADER_NAME_TOKEN.match(stripped) else None + + +def same_header(name: str, other: str) -> bool: + """Whether two HTTP header names are the same one. They are case-insensitive (RFC 7230 3.2).""" + return name.lower() == other.lower() + + +def has_header(headers: Mapping[str, str] | None, name: str) -> bool: + """Whether ``headers`` carries ``name`` under any casing.""" + return bool(headers) and any(same_header(key, name) for key in headers or {}) + + +def without_header(headers: Mapping[str, str] | None, name: str) -> dict[str, str] | None: + """A copy of ``headers`` with every casing of ``name`` removed, or None if nothing remains. + + The one owner of "drop this credential's header". Both MCP stacks and the upstream-credential + resolver share it so a slot can never be dropped case-sensitively in one place and + case-insensitively in another, which is how an injected header came to shadow a resolved + credential on the v1 path. + """ + if not headers: + return None + filtered: Final = {key: value for key, value in headers.items() if not same_header(key, name)} + return filtered or None + + +_DEFAULT_PORTS: Final[Mapping[str, int]] = MappingProxyType({"http": 80, "https": 443}) + + +def crosses_origin(configured: str, target: str) -> bool: + """Whether ``target`` leaves ``configured``'s origin, by the rule HTTP clients use. + + Origin is scheme, host and port, not host alone, so a same-host HTTPS downgrade or a port change + counts as crossing it. A plain http -> https upgrade of the same host is exempt, matching what + httpx exempts when it decides whether to keep ``Authorization`` across a redirect. + """ + a: Final = urlsplit(configured) + b: Final = urlsplit(target) + port_a: Final = a.port or _DEFAULT_PORTS.get(a.scheme) + port_b: Final = b.port or _DEFAULT_PORTS.get(b.scheme) + if a.scheme == b.scheme and a.hostname == b.hostname and port_a == port_b: + return False + return not ( + a.hostname == b.hostname and a.scheme == "http" and port_a == 80 and b.scheme == "https" and port_b == 443 + ) + + +def custom_credential_slot(headers: Mapping[str, str] | None) -> str | None: + """The first header carrying a credential somewhere other than ``Authorization``, if any.""" + return next((name for name in headers or {} if not same_header(name, DEFAULT_CREDENTIAL_HEADER)), None) + + +def credential_redirect_hook( + configured_url: str, slot: str | None +) -> Callable[[httpx.Request], Awaitable[None]] | None: + """An httpx request hook dropping ``slot`` once a redirect leaves ``configured_url``'s origin. + + None when no guard is needed, so callers do not each repeat the exemption: HTTP clients already + strip ``Authorization`` across origins, but forward every other header, so only a credential an + operator moved to its own slot can be replayed to whatever host the upstream redirects to. + """ + if not configured_url or not slot or same_header(slot, DEFAULT_CREDENTIAL_HEADER): + return None + + async def guard(request: httpx.Request) -> None: + if slot in request.headers and crosses_origin(configured_url, str(request.url)): + del request.headers[slot] + + return guard + + +MCP_ADMIN_CONFIG_CREDENTIAL_KEYS: Final[tuple[str, ...]] = ("upstream_resource", "upstream_token_header") """Non-secret credential keys returned on read so the admin form can show and clear them. Mirrors ``ADMIN_CONFIG_CREDENTIAL_KEYS`` in ``ui/litellm-dashboard/src/components/mcp_tools/types.tsx``.""" diff --git a/litellm/types/mcp_server/mcp_server_manager.py b/litellm/types/mcp_server/mcp_server_manager.py index 401793a79e4..9bf3acc601c 100644 --- a/litellm/types/mcp_server/mcp_server_manager.py +++ b/litellm/types/mcp_server/mcp_server_manager.py @@ -1,7 +1,7 @@ from datetime import datetime from typing import Any, Final, Literal -from pydantic import BaseModel, ConfigDict +from pydantic import BaseModel, ConfigDict, field_validator from litellm.types.mcp import ( DEFAULT_SUBJECT_TOKEN_TYPE, @@ -9,6 +9,7 @@ from litellm.types.mcp import ( MCPAuthType, MCPTokenEndpointAuthMethod, MCPTransportType, + normalize_upstream_header_name, ) # MCPInfo now allows arbitrary additional fields for custom metadata @@ -86,6 +87,22 @@ class MCPServer(BaseModel): # today's behavior; "auto" derives the canonical URI from ``url``; any other value is sent # verbatim. Resolved by ``oauth_utils.resolve_upstream_resource``. upstream_resource: str | None = None + # Which upstream header carries the credential LiteLLM resolves for this server (the minted + # OAuth token, or the static key). None keeps RFC 6750's default, ``Authorization``. An ESB or + # API gateway that terminates its own credential in a private header needs this so a second, + # operator-configured ``Authorization`` can pass through to the origin untouched. + upstream_token_header: str | None = None + + @field_validator("upstream_token_header") + @classmethod + def _check_upstream_token_header(cls, value: str | None) -> str | None: + if value is None or not value.strip(): + return None + normalized: Final = normalize_upstream_header_name(value) + if normalized is None: + raise ValueError(f"upstream_token_header must be a valid HTTP header name (RFC 7230 token), got {value!r}") + return normalized + # AWS SigV4 fields aws_access_key_id: str | None = None aws_secret_access_key: str | None = None diff --git a/tests/test_litellm/experimental_mcp_client/test_mcp_client.py b/tests/test_litellm/experimental_mcp_client/test_mcp_client.py index b1182dd7262..fd7ab3afdab 100644 --- a/tests/test_litellm/experimental_mcp_client/test_mcp_client.py +++ b/tests/test_litellm/experimental_mcp_client/test_mcp_client.py @@ -9,6 +9,7 @@ from unittest.mock import AsyncMock, MagicMock, patch import anyio import httpx import pytest +from litellm.proxy._experimental.mcp_server.outbound_credentials.httpx_auth import StaticHeaderAuth from mcp import McpError from mcp.shared.message import SessionMessage from mcp.types import ( @@ -1095,3 +1096,188 @@ def test_mcp_extra_matches_proxy_extra_and_supports_streamable_http(): specifier = Requirement(mcp_extra[0]).specifier assert not specifier.contains("1.23.0") assert specifier.contains("1.28.1") + + +@pytest.mark.parametrize( + "auth_type, default_header", + [ + (MCPAuth.oauth2, "Authorization"), + (MCPAuth.bearer_token, "Authorization"), + (MCPAuth.api_key, "X-API-Key"), + ], +) +def test_v1_auth_headers_default_to_the_auth_type_slot(auth_type: MCPAuth, default_header: str) -> None: + client = MCPClient(server_url="http://up.example.com/mcp", auth_type=auth_type) + client.update_auth_value("tok") + assert default_header in client._get_auth_headers() + + +@pytest.mark.parametrize("auth_type", [MCPAuth.oauth2, MCPAuth.bearer_token, MCPAuth.api_key]) +def test_v1_auth_headers_honor_the_configured_slot(auth_type: MCPAuth) -> None: + """The v1 stack mints its own client_credentials token (oauth2_token_cache) and writes it here, + so leaving this table hardcoded makes the knob a silent no-op for every server that resolves + through v1 rather than the v2 resolver.""" + client = MCPClient( + server_url="http://up.example.com/mcp", + auth_type=auth_type, + auth_header_name="esb-oauth", + ) + client.update_auth_value("tok") + headers = client._get_auth_headers() + assert "esb-oauth" in headers + assert "Authorization" not in headers + assert "X-API-Key" not in headers + + +def test_v1_static_headers_still_win_their_own_slot(): + # extra_headers (which carries static_headers) is applied last on the v1 path, so a static + # Authorization survives untouched while the resolved credential sits on its own header. + client = MCPClient( + server_url="http://up.example.com/mcp", + auth_type=MCPAuth.oauth2, + auth_header_name="esb-oauth", + extra_headers={"Authorization": "Bearer static-upstream-mcp-token"}, + ) + client.update_auth_value("minted") + headers = client._get_auth_headers() + assert headers["esb-oauth"] == "Bearer minted" + assert headers["Authorization"] == "Bearer static-upstream-mcp-token" + + +@pytest.mark.asyncio +async def test_a_custom_credential_header_is_stripped_when_a_redirect_crosses_origin(): + """httpx drops Authorization across origins but keeps every other header, so a credential the + operator moved to its own slot would be replayed to whatever host the upstream redirects to. + Verified against real httpx redirect handling, not a hand-built request. + """ + seen: "list[tuple[str, str]]" = [] + + def handler(request: httpx.Request) -> httpx.Response: + seen.append((request.url.host, request.headers.get("esb-oauth", ""))) + if request.url.host == "upstream.example.com": + return httpx.Response(302, headers={"Location": "https://attacker.example.com/collect"}) + return httpx.Response(200) + + client = MCPClient( + server_url="https://upstream.example.com/mcp", + auth_type=MCPAuth.oauth2, + auth_header_name="esb-oauth", + ) + client.update_auth_value("minted-token") + factory = client._create_httpx_client_factory() + async with factory(headers=client._get_auth_headers(), timeout=None) as http_client: + http_client._transport = httpx.MockTransport(handler) + await http_client.get("https://upstream.example.com/mcp") + + assert seen[0] == ("upstream.example.com", "Bearer minted-token") + assert seen[1] == ("attacker.example.com", "") + + +@pytest.mark.asyncio +async def test_authorization_is_left_to_httpx_and_needs_no_guard(): + # The default slot is already protected by httpx, so the client must not install a guard for it + # and must not interfere with the ordinary Authorization path. + url = "https://upstream.example.com/mcp" + from litellm.types.mcp import credential_redirect_hook + + def guard_for(client: MCPClient): + return credential_redirect_hook(client.server_url, client._credential_slot) + + assert guard_for(MCPClient(server_url=url, auth_type=MCPAuth.oauth2)) is None + assert guard_for(MCPClient(server_url=url, resolved_auth=StaticHeaderAuth("Bearer x"))) is None + # a v2 resolver slot is discovered from the auth object, without the caller naming it again + custom = MCPClient(server_url=url, resolved_auth=StaticHeaderAuth("Bearer x", header_name="esb-oauth")) + assert guard_for(custom) is not None + # and the same answer arrives via the v1 configured slot + assert guard_for(MCPClient(server_url=url, auth_header_name="ESB-OAuth")) is not None + + +def test_an_injected_header_cannot_shadow_the_configured_credential_slot(): + """The v2 path drops a colliding injected header so the resolved credential wins its slot. The + v1 path applies extra_headers last, so without this it silently sends the injected value and the + upstream rejects a credential the gateway thought it had sent. + """ + client = MCPClient( + server_url="https://upstream.example.com/mcp", + auth_type=MCPAuth.oauth2, + auth_header_name="esb-oauth", + extra_headers={"esb-oauth": "Bearer injected", "X-Trace": "keep"}, + ) + client.update_auth_value("minted-token") + headers = client._get_auth_headers() + assert headers["esb-oauth"] == "Bearer minted-token" + assert headers["X-Trace"] == "keep" + + +def test_without_a_configured_slot_the_existing_precedence_is_unchanged(): + # extra_headers winning over authentication_token is long-standing v1 behavior; the fix above + # must apply only to the slot the operator explicitly named. + client = MCPClient( + server_url="https://upstream.example.com/mcp", + auth_type=MCPAuth.oauth2, + extra_headers={"Authorization": "Bearer injected"}, + ) + client.update_auth_value("minted-token") + assert client._get_auth_headers()["Authorization"] == "Bearer injected" + + +_REDIRECT_CASES = [ + ("https://upstream.example.com/mcp", "https://upstream.example.com/other"), # same origin + ("https://upstream.example.com/mcp", "https://upstream.example.com:443/other"), # explicit default port + ("https://upstream.example.com/mcp", "https://attacker.example.com/collect"), # different host + ("https://upstream.example.com/mcp", "http://upstream.example.com/collect"), # scheme downgrade + ("https://upstream.example.com/mcp", "https://upstream.example.com:8443/other"), # different port + ("https://upstream.example.com/mcp", "https://sub.upstream.example.com/x"), # different host + ("http://upstream.example.com/mcp", "https://upstream.example.com/other"), # http -> https upgrade + ("http://upstream.example.com/mcp", "http://upstream.example.com/other"), # same origin, plain http +] + + +@pytest.mark.parametrize("start,target", _REDIRECT_CASES) +@pytest.mark.asyncio +async def test_the_guard_agrees_with_httpx_about_authorization(start: str, target: str) -> None: + """Our custom slot must be dropped on exactly the redirects where httpx drops Authorization. + + The rule is mirrored rather than imported, so this drives real httpx and compares the two + outcomes. A future httpx that changes its redirect rule reds here instead of silently leaving + the custom slot forwarded where Authorization is not (or stripped where it is not needed). + """ + seen: "list[tuple[str, str, str]]" = [] + + def handler(request: httpx.Request) -> httpx.Response: + seen.append( + ( + str(request.url), + request.headers.get("authorization", ""), + request.headers.get("esb-oauth", ""), + ) + ) + if str(request.url) == start: + return httpx.Response(302, headers={"Location": target}) + return httpx.Response(200) + + client = MCPClient(server_url=start, auth_type=MCPAuth.oauth2, auth_header_name="esb-oauth") + factory = client._create_httpx_client_factory() + async with factory(headers={"Authorization": "Bearer AUTH", "esb-oauth": "Bearer ESB"}, timeout=None) as http: + http._transport = httpx.MockTransport(handler) + await http.get(start) + + _url, authorization, esb = seen[-1] + assert (authorization == "") == (esb == ""), ( + f"httpx and the guard disagree for {target}: authorization={authorization!r} esb-oauth={esb!r}" + ) + + +def test_a_differently_cased_injected_header_cannot_shadow_the_slot() -> None: + # HTTP header names are case-insensitive and v2 drops the collision case-insensitively, so an + # exact-key check here would leave both spellings in the dict and let the injected value win. + client = MCPClient( + server_url="https://upstream.example.com/mcp", + auth_type=MCPAuth.oauth2, + auth_header_name="esb-oauth", + extra_headers={"ESB-OAuth": "Bearer injected", "X-Trace": "keep"}, + ) + client.update_auth_value("minted-token") + headers = client._get_auth_headers() + assert [v for k, v in headers.items() if k.lower() == "esb-oauth"] == ["Bearer minted-token"] + assert headers["X-Trace"] == "keep" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py index 0020dbf8d61..c667db7f07c 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py @@ -10,6 +10,7 @@ from types import SimpleNamespace import pytest from fastapi import HTTPException +from pydantic import ValidationError from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import ( oauth_protected_resource_path, @@ -598,3 +599,92 @@ def test_client_credentials_uses_admin_entered_token_url_when_issuer_yield_empti assert spec is not None assert isinstance(spec.config, ClientCredentialsConfig) assert spec.config.token_url == "https://idp.example.com/token" + + +_M2M_FIELDS = dict( + auth_type=MCPAuth.oauth2, + oauth2_flow="client_credentials", + client_id="cid", + client_secret="csec", + token_url="https://idp.example.com/token", +) +_OBO_FIELDS = dict( + auth_type=MCPAuth.oauth2_token_exchange, + client_id="cid", + client_secret="csec", + token_exchange_endpoint="https://idp.example.com/token", +) +_ID_JAG_FIELDS = dict( + auth_type=MCPAuth.oauth2_id_jag, + client_id="cid", + client_secret="csec", + token_exchange_endpoint="https://idp.example.com/token", + id_jag_resource_token_endpoint="https://mcp-as.example.com/token", + audience="api://mcp", +) +_AUTHZ_CODE_FIELDS = dict(auth_type=MCPAuth.oauth2, url="https://up.example.com/mcp") +_STATIC_FIELDS = dict(auth_type=MCPAuth.bearer_token, authentication_token="static-tok") + +_ARM_FIELDS = ( + ("client_credentials", _M2M_FIELDS), + ("token_exchange", _OBO_FIELDS), + ("id_jag", _ID_JAG_FIELDS), + ("authorization_code", _AUTHZ_CODE_FIELDS), + ("api_key", _STATIC_FIELDS), +) + + +@pytest.mark.parametrize("name,fields", _ARM_FIELDS, ids=[n for n, _ in _ARM_FIELDS]) +def test_upstream_token_header_reaches_every_arms_config(name, fields): + # to_server_spec builds each arm's config from a hand-written kwargs list, so an arm that + # forgets to read the field fails silently: the server keeps writing to Authorization. + spec = to_server_spec(_server(upstream_token_header="esb-oauth", **fields)) + assert spec is not None + assert spec.config.header_name == "esb-oauth" + + +@pytest.mark.parametrize("name,fields", _ARM_FIELDS, ids=[n for n, _ in _ARM_FIELDS]) +def test_omitting_the_field_keeps_each_arms_shipped_default(name, fields): + spec = to_server_spec(_server(**fields)) + assert spec is not None + assert spec.config.header_name == "Authorization" + + +def test_api_key_scheme_default_survives_when_the_field_is_unset(): + spec = to_server_spec(_server(auth_type=MCPAuth.api_key, authentication_token="k")) + assert spec is not None + assert spec.config.header_name == "X-API-Key" + assert spec.config.value_prefix == "" + + +def test_the_field_overrides_the_api_key_scheme_default(): + spec = to_server_spec(_server(auth_type=MCPAuth.api_key, authentication_token="k", upstream_token_header="X-Esb")) + assert spec is not None + assert spec.config.header_name == "X-Esb" + + +@pytest.mark.parametrize("bad", ["with space", "has:colon", "trailing\r\nX-Injected", 'quoted"name']) +def test_a_malformed_header_name_is_refused_when_the_server_is_built(bad): + """Validation belongs at ingestion, not at spec building. Raising inside to_server_spec would + abort the whole aggregate tools/list, so one mistyped server would silently empty the tool list + for every other server too. Refusing at MCPServer construction fails the config load loudly + instead, and means no malformed value can ever reach an arm. + """ + with pytest.raises(ValidationError): + _server(upstream_token_header=bad, **_M2M_FIELDS) + + +def test_a_valid_header_name_is_trimmed_at_ingestion(): + assert _server(upstream_token_header=" esb-oauth ", **_M2M_FIELDS).upstream_token_header == "esb-oauth" + + +@pytest.mark.parametrize("blank", ["", " ", "\t"]) +def test_a_blank_header_name_means_unset_rather_than_an_error(blank): + """The management API treats a blank as "not supplied" and stores it, so raising here made every + later rebuild of that server 500 instead of falling back to the default Authorization behavior. + """ + server = _server(upstream_token_header=blank, **_M2M_FIELDS) + assert server.upstream_token_header is None + spec = to_server_spec(server) + assert spec is not None + assert spec.config.header_name == "Authorization" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_client_credentials.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_client_credentials.py index a5d17428b37..010e7e14d39 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_client_credentials.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_client_credentials.py @@ -341,7 +341,7 @@ async def test_bearer_auth_sends_the_token_and_leaves_a_success_alone(): async def refetch(failed: str) -> "str | None": raise AssertionError("must not refetch on success") - auth = ClientCredentialsBearerAuth("m2m-token", refetch) + auth = ClientCredentialsBearerAuth("m2m-token", refetch, ClientCredentialsConfig()) async with httpx.AsyncClient(transport=transport, auth=auth) as client: response = await client.get("https://upstream.example.com/mcp") assert response.status_code == 200 @@ -357,7 +357,7 @@ async def test_bearer_auth_retries_a_401_once_with_a_fresh_token(): refetched.append(failed) return "fresh-token" - auth = ClientCredentialsBearerAuth("stale-token", refetch) + auth = ClientCredentialsBearerAuth("stale-token", refetch, ClientCredentialsConfig()) async with httpx.AsyncClient(transport=transport, auth=auth) as client: response = await client.get("https://upstream.example.com/mcp") assert response.status_code == 200 @@ -377,7 +377,7 @@ async def test_bearer_auth_remembers_the_rotated_token_for_later_requests(): refetched.append(failed) return "fresh-token" - auth = ClientCredentialsBearerAuth("stale-token", refetch) + auth = ClientCredentialsBearerAuth("stale-token", refetch, ClientCredentialsConfig()) async with httpx.AsyncClient(transport=transport, auth=auth) as client: first = await client.get("https://upstream.example.com/mcp") second = await client.get("https://upstream.example.com/mcp") @@ -393,7 +393,7 @@ async def test_bearer_auth_surfaces_the_401_when_the_refetch_fails(): async def refetch(failed: str) -> "str | None": return None - auth = ClientCredentialsBearerAuth("stale-token", refetch) + auth = ClientCredentialsBearerAuth("stale-token", refetch, ClientCredentialsConfig()) async with httpx.AsyncClient(transport=transport, auth=auth) as client: response = await client.get("https://upstream.example.com/mcp") assert response.status_code == 401 @@ -409,7 +409,7 @@ async def test_bearer_auth_gives_up_after_a_second_401(): refetched.append(failed) return "fresh-token" - auth = ClientCredentialsBearerAuth("stale-token", refetch) + auth = ClientCredentialsBearerAuth("stale-token", refetch, ClientCredentialsConfig()) async with httpx.AsyncClient(transport=transport, auth=auth) as client: response = await client.get("https://upstream.example.com/mcp") assert response.status_code == 401 @@ -421,7 +421,60 @@ def test_bearer_auth_rejects_sync_clients(): async def refetch(failed: str) -> "str | None": return None - auth = ClientCredentialsBearerAuth("token", refetch) + auth = ClientCredentialsBearerAuth("token", refetch, ClientCredentialsConfig()) with httpx.Client(transport=httpx.MockTransport(lambda request: httpx.Response(200)), auth=auth) as client: with pytest.raises(RuntimeError): client.get("https://upstream.example.com/mcp") + + +@pytest.mark.asyncio +async def test_bearer_auth_writes_the_minted_token_to_the_configured_header(): + seen: "list[dict[str, str]]" = [] + + def handler(request: httpx.Request) -> httpx.Response: + seen.append(dict(request.headers)) + return httpx.Response(200) + + async def refetch(failed: str) -> "str | None": + raise AssertionError("must not refetch on success") + + auth = ClientCredentialsBearerAuth("m2m-token", refetch, ClientCredentialsConfig(header_name="esb-oauth")) + async with httpx.AsyncClient(transport=httpx.MockTransport(handler), auth=auth) as client: + await client.get("https://upstream.example.com/mcp") + assert seen[0]["esb-oauth"] == "Bearer m2m-token" + assert "authorization" not in seen[0] + + +@pytest.mark.asyncio +async def test_the_401_refetch_retry_also_targets_the_configured_header(): + # The retry is a SECOND write of the credential. Honoring the carrier only on the first write + # would silently send the fresh token to Authorization, so the ESB rejects every recovered + # request while the first attempt looked correct. + seen: "list[dict[str, str]]" = [] + responses = [httpx.Response(401), httpx.Response(200)] + + def handler(request: httpx.Request) -> httpx.Response: + seen.append(dict(request.headers)) + return responses[min(len(seen) - 1, len(responses) - 1)] + + async def refetch(failed: str) -> "str | None": + return "fresh-token" + + auth = ClientCredentialsBearerAuth("stale-token", refetch, ClientCredentialsConfig(header_name="esb-oauth")) + async with httpx.AsyncClient(transport=httpx.MockTransport(handler), auth=auth) as client: + response = await client.get("https://upstream.example.com/mcp") + assert response.status_code == 200 + assert [h["esb-oauth"] for h in seen] == ["Bearer stale-token", "Bearer fresh-token"] + assert all("authorization" not in h for h in seen) + + +@pytest.mark.asyncio +async def test_bearer_auth_advertises_the_header_it_will_occupy(): + # _resolve_v2_auth reads header_name off the auth object to decide which injected header + # conflicts; an auth object that lies about its slot would drop the wrong one. + async def refetch(failed: str) -> "str | None": + return None + + assert ClientCredentialsBearerAuth("t", refetch, ClientCredentialsConfig()).header_name == "Authorization" + default_carrier = ClientCredentialsConfig(header_name="esb-oauth") + assert ClientCredentialsBearerAuth("t", refetch, default_carrier).header_name == "esb-oauth" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py index 0d130767bd5..9d63e8c2c1c 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py @@ -1033,3 +1033,70 @@ async def test_invalidate_credentials_for_id_jag_is_a_noop_without_a_caller_toke assert isinstance(first, Ok) and isinstance(second, Ok) assert _emitted(second.ok)["Authorization"] == "Bearer cached-bearer" assert len(endpoint.calls) == 2 + + +async def _resolve_with_carrier(kind: str, header: str): + """Resolve one minted-token arm whose config targets ``header``.""" + if kind == "client_credentials": + source = _FakeM2MSource(Ok(OAuthToken(access_token="minted"))) + config = _M2M.model_copy(update={"header_name": header}) + provider = UpstreamCredentialProvider(client_credentials_source=source) + return await provider.resolve_credentials(_SUBJECT, _spec(config)) + if kind == "token_exchange": + exchanger = _FakeExchanger(Ok(OAuthToken(access_token="minted"))) + config = _OBO.model_copy(update={"header_name": header}) + subject = Subject(tenant_id="acme", subject_id="alice", inbound_token=SecretStr("caller-jwt")) + provider = UpstreamCredentialProvider(token_exchanger=exchanger) + return await provider.resolve_credentials(subject, _spec(config)) + if kind == "authorization_code": + store = _FakeTokenStore({("alice", "s"): OAuthToken(access_token="minted")}) + provider = UpstreamCredentialProvider(oauth_token_store=store) + return await provider.resolve_credentials( + Subject(tenant_id="", subject_id="alice"), + _spec(AuthorizationCodeConfig(header_name=header)), + ) + endpoint = _FakeTokenEndpoint( + [ + Ok(ExchangedToken(access_token="id-jag-assertion", expires_in=300)), + Ok(ExchangedToken(access_token="minted", expires_in=300)), + ] + ) + config = _id_jag_config().model_copy(update={"header_name": header}) + subject = Subject(tenant_id="acme", subject_id="alice", inbound_token=SecretStr("caller-id-token")) + provider = UpstreamCredentialProvider(token_endpoint=endpoint) + return await provider.resolve_credentials(subject, _spec(config)) + + +_MINTED_ARMS = ("client_credentials", "token_exchange", "authorization_code", "id_jag") + + +@pytest.mark.parametrize("kind", _MINTED_ARMS) +@pytest.mark.asyncio +async def test_every_minted_arm_emits_its_configured_header(kind): + # One arm left on a hardcoded Authorization is a silent no-op for exactly the server that + # configured the knob, so this is asserted across all four rather than on the M2M arm alone. + result = await _resolve_with_carrier(kind, "esb-oauth") + assert isinstance(result, Ok) + headers, _ = await _emitted_async(result.ok) + assert headers["esb-oauth"] == "Bearer minted" + assert "authorization" not in headers + + +@pytest.mark.parametrize("kind", _MINTED_ARMS) +@pytest.mark.asyncio +async def test_every_minted_arm_still_defaults_to_authorization(kind): + result = await _resolve_with_carrier(kind, "Authorization") + assert isinstance(result, Ok) + headers, _ = await _emitted_async(result.ok) + assert headers["Authorization"] == "Bearer minted" + + +@pytest.mark.asyncio +async def test_passthrough_ignores_the_carrier_and_keeps_the_callers_slot(): + # Passthrough mints nothing: it forwards the caller's own credential, so it has no carrier to + # configure and must keep using the header the caller aimed it at. + subject = Subject(tenant_id="", subject_id="", inbound_token=SecretStr("caller-token")) + result = await UpstreamCredentialProvider().resolve_credentials(subject, _spec(PassthroughConfig())) + assert isinstance(result, Ok) + headers, _ = await _emitted_async(result.ok) + assert headers["Authorization"] == "caller-token" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_types.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_types.py index bb25ab6bd3c..d4b51b08e06 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_types.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_types.py @@ -14,9 +14,11 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials import ( Ambient, ApiKeyConfig, AuthConfig, + AuthorizationCodeConfig, AuthSpecKind, AwsSigV4Config, Byok, + ClientCredentialsConfig, ClientSecretAuth, CredError, Error, @@ -27,7 +29,9 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials import ( ServerSpec, SharedKey, StaticKeys, + TokenExchangeConfig, parse_auth_spec_kind, + validate_header_name, ) _AUTH_CONFIG = TypeAdapter(AuthConfig) @@ -229,3 +233,61 @@ def test_id_jag_server_spec_derives_auth_spec_kind(): config=config, ) assert spec.auth_spec_kind is AuthSpecKind.id_jag + + +_CARRIER_CONFIGS = ( + ("client_credentials", ClientCredentialsConfig), + ("token_exchange", lambda **kw: TokenExchangeConfig(token_exchange_endpoint="https://idp/te", **kw)), + ("authorization_code", AuthorizationCodeConfig), + ( + "id_jag", + lambda **kw: IdJagConfig( + org_token_endpoint="https://idp.example.com/token", + resource_token_endpoint="https://mcp-as.example.com/token", + client_id="litellm", + client_auth=ClientSecretAuth(client_secret=SecretStr("s")), + **kw, + ), + ), + ("api_key", lambda **kw: ApiKeyConfig(key_source=SharedKey(value=SecretStr("k")), **kw)), +) + + +@pytest.mark.parametrize("name,build", _CARRIER_CONFIGS, ids=[n for n, _ in _CARRIER_CONFIGS]) +def test_every_resolved_credential_config_defaults_to_rfc6750_authorization(name, build): + # The default is what preserves today's wire behavior for every existing server. + assert build().header("tok") == ("Authorization", "Bearer tok") + + +@pytest.mark.parametrize("name,build", _CARRIER_CONFIGS, ids=[n for n, _ in _CARRIER_CONFIGS]) +def test_every_resolved_credential_config_honors_a_custom_header(name, build): + assert build(header_name="esb-oauth").header("tok") == ("esb-oauth", "Bearer tok") + + +@pytest.mark.parametrize("name,build", _CARRIER_CONFIGS, ids=[n for n, _ in _CARRIER_CONFIGS]) +def test_every_resolved_credential_config_can_send_a_raw_value(name, build): + assert build(header_name="esb-oauth", value_prefix="").header("tok") == ("esb-oauth", "tok") + + +@pytest.mark.parametrize( + "bad", + [ + "with space", + "has:colon", + "trailing\r\nX-Injected", + "", + " ", + "quoted\"name", + ], +) +def test_header_name_outside_the_rfc7230_token_grammar_is_rejected(bad): + # An operator-supplied name reaches egress verbatim, so anything that could split a + # header must fail closed at construction rather than be sanitized later. + with pytest.raises(ValidationError): + ClientCredentialsConfig(header_name=bad) + assert isinstance(validate_header_name(bad), Error) + + +def test_header_name_is_trimmed_by_the_one_validator(): + assert validate_header_name(" esb-oauth ") == Ok("esb-oauth") + assert ClientCredentialsConfig(header_name=" esb-oauth ").header_name == "esb-oauth" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py index cdea803ebf3..5508259273d 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py @@ -43,7 +43,6 @@ from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( _obo_retry_applies, _resolve_openapi_tool_auth, _should_strip_caller_authorization, - _without_authorization, ) from litellm.proxy._types import ( LiteLLM_MCPServerTable, @@ -2405,6 +2404,104 @@ class TestMCPServerManager: assert client._resolved_auth is not None assert "authorization" not in {k.lower() for k in (client.extra_headers or {})} + @staticmethod + def _esb_server(header: "str | None") -> MCPServer: + return MCPServer( + server_id="esb", + name="esb-server", + url="https://up.example.com/mcp", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + oauth2_flow="client_credentials", + client_id="cid", + client_secret="csec", + token_url="https://idp.example.com/token", + upstream_token_header=header, + static_headers={"Authorization": "Bearer static-upstream-mcp-token"}, + ) + + @pytest.mark.asyncio + async def test_static_authorization_survives_a_minted_token_aimed_elsewhere(self): + """The dual-credential case: an ESB wants the gateway-minted token on its own header while a + separate static Authorization passes through to the origin. Dropping Authorization here (the + old name-blind behavior) deletes the second credential and the upstream 401s.""" + from litellm.proxy._experimental.mcp_server.outbound_credentials.httpx_auth import ( + StaticHeaderAuth, + ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.result import Ok + + class _FakeProvider: + async def resolve_credentials(self, subject, server): + return Ok(StaticHeaderAuth("Bearer MINTED-M2M", header_name="esb-oauth")) + + manager = MCPServerManager(cred_provider=_FakeProvider()) + client = await manager._create_mcp_client( + self._esb_server("esb-oauth"), + extra_headers={"Authorization": "Bearer static-upstream-mcp-token"}, + ) + + assert client._resolved_auth is not None + assert (client.extra_headers or {})["Authorization"] == "Bearer static-upstream-mcp-token" + + @pytest.mark.asyncio + async def test_a_minted_token_aimed_at_the_static_header_still_wins_that_slot(self): + """The negative class of the test above: when the two DO collide the resolver-owned + credential is still authoritative, so the knob cannot be used to smuggle a second + credential into the same slot.""" + from litellm.proxy._experimental.mcp_server.outbound_credentials.httpx_auth import ( + StaticHeaderAuth, + ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.result import Ok + + class _FakeProvider: + async def resolve_credentials(self, subject, server): + return Ok(StaticHeaderAuth("Bearer MINTED-M2M", header_name="esb-oauth")) + + manager = MCPServerManager(cred_provider=_FakeProvider()) + client = await manager._create_mcp_client( + self._esb_server("esb-oauth"), + extra_headers={"esb-oauth": "Bearer signer-jwt", "X-Trace": "keep-me"}, + ) + + assert client._resolved_auth is not None + assert "esb-oauth" not in {k.lower() for k in (client.extra_headers or {})} + assert (client.extra_headers or {})["X-Trace"] == "keep-me" + + @pytest.mark.asyncio + async def test_a_differently_cased_injected_header_is_still_recognised_as_the_collision(self): + """HTTP header names are case-insensitive, so the conflict check must be too. + + A case-sensitive check reports no conflict and hands the injected header back untouched, so + the returned extra_headers still carries a second copy of the credential slot for every + downstream consumer of that dict. httpx happens to collapse the two on the wire, which is + exactly why this needs pinning rather than being left to luck. + """ + from litellm.proxy._experimental.mcp_server.outbound_credentials.httpx_auth import ( + StaticHeaderAuth, + ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.result import Ok + + class _FakeProvider: + async def resolve_credentials(self, subject, server): + return Ok(StaticHeaderAuth("Bearer MINTED", header_name="esb-oauth")) + + manager = MCPServerManager(cred_provider=_FakeProvider()) + client = await manager._create_mcp_client( + self._esb_server("esb-oauth"), + extra_headers={"ESB-OAuth": "Bearer injected", "X-Trace": "keep"}, + ) + + assert client._resolved_auth is not None + assert not any(k.lower() == "esb-oauth" for k in (client.extra_headers or {})) + assert (client.extra_headers or {})["X-Trace"] == "keep" + + def test_without_header_drops_only_the_named_header(self): + from litellm.types.mcp import DEFAULT_CREDENTIAL_HEADER, without_header + + headers = {"Authorization": "Bearer a", "esb-oauth": "Bearer b", "X-Trace": "t"} + assert without_header(headers, "ESB-OAuth") == {"Authorization": "Bearer a", "X-Trace": "t"} + assert without_header(headers, DEFAULT_CREDENTIAL_HEADER) == {"esb-oauth": "Bearer b", "X-Trace": "t"} + @pytest.mark.asyncio async def test_preflight_token_exchange_challenges_on_rejected_subject(self): """A subject the IdP rejects must raise the RFC 9728 401 challenge from the preflight, so a @@ -2624,14 +2721,16 @@ class TestMCPServerManager: if captured_extra_headers: assert "authorization" not in {k.lower() for k in captured_extra_headers} - def test_without_authorization_drops_only_the_credential(self): + def test_without_header_drops_only_the_credential(self): + from litellm.types.mcp import without_header + # None / empty -> None - assert _without_authorization(None) is None - assert _without_authorization({}) is None + assert without_header(None, "Authorization") is None + assert without_header({}, "Authorization") is None # Only Authorization present -> nothing left -> None (case-insensitive) - assert _without_authorization({"authorization": "Bearer x"}) is None + assert without_header({"authorization": "Bearer x"}, "Authorization") is None # Authorization dropped, other headers kept - assert _without_authorization({"Authorization": "Bearer x", "X-Trace-Id": "t"}) == {"X-Trace-Id": "t"} + assert without_header({"Authorization": "Bearer x", "X-Trace-Id": "t"}, "Authorization") == {"X-Trace-Id": "t"} @pytest.mark.asyncio async def test_call_regular_mcp_tool_passthrough_forwards_authorization_with_admission_header( @@ -9641,13 +9740,38 @@ class TestMaterializeAuthHeaders: from litellm.proxy._experimental.mcp_server.outbound_credentials.client_credentials import ( ClientCredentialsBearerAuth, ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( + ClientCredentialsConfig, + ) async def _refetch(_stale: str): return None - headers = await _materialize_auth_headers(ClientCredentialsBearerAuth("m2m-token", _refetch)) + default_carrier = ClientCredentialsConfig() + headers = await _materialize_auth_headers(ClientCredentialsBearerAuth("m2m-token", _refetch, default_carrier)) assert headers == {"Authorization": "Bearer m2m-token"} + @pytest.mark.asyncio + async def test_materialize_follows_the_minted_token_to_a_custom_header(self): + # The OpenAPI arm reads header_name off the auth object rather than assuming Authorization, + # so it carries the knob with no per-arm change. This pins that it stays that way. + from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( + _materialize_auth_headers, + ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.client_credentials import ( + ClientCredentialsBearerAuth, + ) + from litellm.proxy._experimental.mcp_server.outbound_credentials.types import ( + ClientCredentialsConfig, + ) + + async def _refetch(_stale: str): + return None + + esb_carrier = ClientCredentialsConfig(header_name="esb-oauth") + headers = await _materialize_auth_headers(ClientCredentialsBearerAuth("m2m-token", _refetch, esb_carrier)) + assert headers == {"esb-oauth": "Bearer m2m-token"} + @pytest.mark.asyncio async def test_noop_and_none_materialize_to_none(self): from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py index b1aa16a30c0..f7567efcabc 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py @@ -13,6 +13,7 @@ import pytest from litellm.proxy._experimental.mcp_server.oauth2_token_cache import ( MCPOAuth2TokenCache, resolve_mcp_auth, + resolved_token_header, ) from litellm.proxy._types import MCPTransport from litellm.types.mcp import MCPAuth @@ -411,3 +412,50 @@ async def test_m2m_mint_uses_admin_entered_token_url_when_issuer_yield_empties_r assert result == "m2m-token-configured" assert mock_client.post.call_args[0][0] == "https://auth.example.com/token" + + +def _m2m_server(**overrides): + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + fields = dict( + server_id="s", + name="n", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + oauth2_flow="client_credentials", + client_id="cid", + client_secret="csec", + token_url="https://idp.example.com/token", + ) + fields.update(overrides) + return MCPServer(**fields) + + +def test_resolved_token_header_follows_the_configured_header_for_a_gateway_resolved_token(): + # resolve_mcp_auth mints the M2M token on this branch, so the value is the gateway's own and + # follows upstream_token_header. + assert resolved_token_header(_m2m_server(upstream_token_header="esb-oauth")) == "esb-oauth" + + +def test_resolved_token_header_is_none_when_the_server_configures_nothing(): + assert resolved_token_header(_m2m_server()) is None + + +def test_a_caller_supplied_credential_never_moves(): + # The caller aimed their own token at the slot the upstream normally uses. Relocating it would + # break every existing x-mcp-auth caller on a server that sets the field for its own token. + server = _m2m_server(upstream_token_header="esb-oauth") + assert resolved_token_header(server, "Bearer caller-token") is None + assert resolved_token_header(server, {"Authorization": "Bearer caller-token"}) is None + + +def test_the_header_and_the_value_agree_on_which_branch_they_took(): + # The two helpers are read as a pair at one call site, so they must never disagree about + # whether the credential came from the caller or from the server's own config. + import asyncio + + server = _m2m_server(upstream_token_header="esb-oauth", authentication_token="static-tok") + caller = "Bearer caller-token" + assert asyncio.run(resolve_mcp_auth(server, caller)) == caller + assert resolved_token_header(server, caller) is None diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_openapi_tool_auth.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_openapi_tool_auth.py index bd953dc55f3..64614c094ba 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_openapi_tool_auth.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_openapi_tool_auth.py @@ -729,3 +729,121 @@ async def test_local_dispatch_reports_the_outcome_instead_of_success(failure: st # A non-auth upstream failure stays a 200 with isError, so REST does not report it as a gateway 500 assert result.isError is True assert "upstream returned HTTP 429" in result.content[0].text + + +@pytest.mark.parametrize( + "resolved,expect_guard", + [ + ({"esb-oauth": "Bearer minted"}, True), + ({"Authorization": "Bearer minted"}, False), + ({}, False), + ], +) +def test_only_a_custom_credential_slot_needs_the_redirect_guard(resolved, expect_guard): + """The OpenAPI arm sends resolved credentials through a redirect-following client, so a custom + slot needs the same cross-origin guard the MCP client installs. Authorization does not: the HTTP + client already strips that one, and taking the guarded path would give up the shared client. + """ + from litellm.types.mcp import DEFAULT_CREDENTIAL_HEADER, same_header + + guarded = next((n for n in resolved if not same_header(n, DEFAULT_CREDENTIAL_HEADER)), None) + assert (guarded is not None) is expect_guard + + +@pytest.mark.asyncio +async def test_the_openapi_arm_drops_a_custom_slot_across_origins(): + """End to end on the hook the OpenAPI arm installs: same origin keeps the credential, a redirect + to another host does not carry it. + """ + import httpx + + from litellm.types.mcp import credential_redirect_hook + + hook = credential_redirect_hook("https://api.example.com/v1/things", "esb-oauth") + + same = httpx.Request("POST", "https://api.example.com/v1/other", headers={"esb-oauth": "Bearer m"}) + await hook(same) + assert same.headers["esb-oauth"] == "Bearer m" + + foreign = httpx.Request("POST", "https://attacker.example.com/collect", headers={"esb-oauth": "Bearer m"}) + await hook(foreign) + assert "esb-oauth" not in foreign.headers + + +def test_the_openapi_arm_installs_the_guard_when_a_credential_rides_a_custom_slot(): + """Pins the wiring, not just the hook: the arm must actually build a guarded client. Testing the + hook alone passes even if this arm never installs it. + """ + from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import ( + _request_resolved_auth_headers, + _upstream_client, + ) + + token = _request_resolved_auth_headers.set({"esb-oauth": "Bearer minted"}) + try: + client = _upstream_client() + assert client.client.event_hooks["request"], "custom slot must install a redirect guard" + finally: + _request_resolved_auth_headers.reset(token) + + +def test_the_guarded_client_is_reused_rather_than_built_per_call(): + """A fresh handler per guarded call is never closed, so every OpenAPI tool call on a server that + sets upstream_token_header would leak an httpx client and its connection pool. Both variants + have to come from the shared cache. + """ + from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import ( + _request_resolved_auth_headers, + _upstream_client, + ) + + token = _request_resolved_auth_headers.set({"esb-oauth": "Bearer minted"}) + try: + assert _upstream_client() is _upstream_client() + finally: + _request_resolved_auth_headers.reset(token) + + +@pytest.mark.asyncio +async def test_the_shared_guard_reads_the_url_from_the_request_context(): + """The hook is one stable object so the client stays cacheable, which means the origin it guards + against has to arrive per request rather than being closed over. + """ + import httpx + + from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import ( + _drop_credential_across_origin, + _request_resolved_auth_headers, + _request_upstream_url, + ) + + creds = _request_resolved_auth_headers.set({"esb-oauth": "Bearer minted"}) + url = _request_upstream_url.set("https://api.example.com/v1/things") + try: + same = httpx.Request("POST", "https://api.example.com/v1/other", headers={"esb-oauth": "Bearer m"}) + await _drop_credential_across_origin(same) + assert same.headers["esb-oauth"] == "Bearer m" + + foreign = httpx.Request("POST", "https://attacker.example.com/x", headers={"esb-oauth": "Bearer m"}) + await _drop_credential_across_origin(foreign) + assert "esb-oauth" not in foreign.headers + finally: + _request_upstream_url.reset(url) + _request_resolved_auth_headers.reset(creds) + + +@pytest.mark.parametrize("resolved", [{"Authorization": "Bearer minted"}, {}, None]) +def test_the_openapi_arm_keeps_the_shared_client_when_no_guard_is_needed(resolved): + # Authorization is already stripped across origins by the HTTP client, so taking the guarded + # path for it would give up the shared connection pool for nothing. + from litellm.proxy._experimental.mcp_server.openapi_to_mcp_generator import ( + _request_resolved_auth_headers, + _upstream_client, + ) + + token = _request_resolved_auth_headers.set(resolved) + try: + client = _upstream_client() + assert not client.client.event_hooks.get("request") + finally: + _request_resolved_auth_headers.reset(token) diff --git a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py index 0d639e1cb6a..ceb44de5576 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_mcp_management_endpoints.py @@ -1573,6 +1573,7 @@ class TestTemporaryMCPSessionEndpoints: existing_server.aws_region_name = None existing_server.aws_service_name = None existing_server.upstream_resource = None + existing_server.upstream_token_header = None mock_manager = MagicMock() mock_manager.get_mcp_server_by_id.return_value = existing_server @@ -1608,6 +1609,7 @@ class TestTemporaryMCPSessionEndpoints: existing_server.aws_region_name = None existing_server.aws_service_name = None existing_server.upstream_resource = None + existing_server.upstream_token_header = None for key, value in server_overrides.items(): setattr(existing_server, key, value) @@ -1639,6 +1641,23 @@ class TestTemporaryMCPSessionEndpoints: assert updated.credentials["client_id"] == "client-123" assert updated.credentials["client_secret"] == "secret-xyz" + def test_upstream_token_header_is_inherited_like_other_admin_config(self): + """It is admin config rather than a credential, so a session server derived from an existing + one must carry it. Miss it and the derived server silently sends its token to Authorization + while the original sends it to the gateway's header.""" + updated = self._inherit_with({}, upstream_token_header="esb-oauth") + + assert updated.credentials["upstream_token_header"] == "esb-oauth" + + def test_a_supplied_upstream_token_header_does_not_read_as_a_credential(self): + """It is in the admin-config key set, so submitting only it must still inherit the declared + app rather than reading as "the caller supplied real credentials".""" + updated = self._inherit_with({"upstream_token_header": "esb-oauth"}) + + assert updated.credentials["client_id"] == "client-123" + assert updated.credentials["client_secret"] == "secret-xyz" + assert updated.credentials["upstream_token_header"] == "esb-oauth" + def test_supplied_credential_still_wins_over_inheritance(self): """A caller that supplies a real credential keeps it; inheritance must not overwrite it.""" updated = self._inherit_with({"auth_value": "caller-token"}) @@ -2256,6 +2275,7 @@ class TestTemporaryMCPSessionEndpoints: aws_region_name=None, aws_service_name=None, upstream_resource=None, + upstream_token_header=None, ) built_server = generate_mock_mcp_server_config_record(server_id="temp-server") mock_manager = MagicMock() diff --git a/tests/test_litellm/types/test_mcp.py b/tests/test_litellm/types/test_mcp.py new file mode 100644 index 00000000000..5450ec4aa48 --- /dev/null +++ b/tests/test_litellm/types/test_mcp.py @@ -0,0 +1,87 @@ +"""Tests for the shared MCP header primitives. + +``same_header`` / ``has_header`` / ``without_header`` are the one owner of "is this the credential's +header", used by both MCP stacks and the upstream-credential resolver. They live here rather than in +either stack because a second implementation is exactly how an injected header came to shadow a +resolved credential on one path and not the other. +""" + +import pytest + +from litellm.types.mcp import ( + credential_redirect_hook, + crosses_origin, + has_header, + same_header, + without_header, +) + + +@pytest.mark.parametrize( + "a,b,expected", + [ + ("Authorization", "authorization", True), + ("ESB-OAuth", "esb-oauth", True), + ("esb-oauth", "esb-oauth", True), + ("esb-oauth", "esb_oauth", False), + ("esb-oauth", "Authorization", False), + ], +) +def test_header_names_compare_case_insensitively(a: str, b: str, expected: bool) -> None: + # RFC 7230 3.2. Every consumer of a credential slot routes through this, so a case-sensitive + # comparison anywhere would let an injected header shadow a resolved credential. + assert same_header(a, b) is expected + + +def test_without_header_drops_every_casing_and_keeps_the_rest() -> None: + headers = {"ESB-OAuth": "injected", "esb-oauth": "also injected", "X-Trace": "keep"} + assert without_header(headers, "esb-oauth") == {"X-Trace": "keep"} + + +def test_without_header_collapses_to_none_when_nothing_remains() -> None: + assert without_header({"Authorization": "Bearer x"}, "AUTHORIZATION") is None + assert without_header(None, "esb-oauth") is None + assert without_header({}, "esb-oauth") is None + + +def test_has_header_matches_any_casing() -> None: + assert has_header({"ESB-OAuth": "v"}, "esb-oauth") is True + assert has_header({"X-Other": "v"}, "esb-oauth") is False + assert has_header(None, "esb-oauth") is False + + +@pytest.mark.parametrize( + "target,expected", + [ + ("https://upstream.example.com/other", False), # same origin + ("https://upstream.example.com:443/other", False), # explicit default port + ("https://attacker.example.com/collect", True), # different host + ("http://upstream.example.com/collect", True), # scheme downgrade, same host + ("https://upstream.example.com:8443/other", True), # different port, same host + ("https://sub.upstream.example.com/x", True), # different host + ], +) +def test_origin_is_scheme_host_and_port_not_host_alone(target: str, expected: bool) -> None: + assert crosses_origin("https://upstream.example.com/mcp", target) is expected + + +def test_an_https_upgrade_of_the_same_host_is_not_crossing() -> None: + # HTTP clients exempt this when deciding to keep Authorization, so a credential slot that did + # not would lose the credential on every such redirect. + assert crosses_origin("http://upstream.example.com/mcp", "https://upstream.example.com/x") is False + assert crosses_origin("http://upstream.example.com/mcp", "http://upstream.example.com/x") is False + + +@pytest.mark.asyncio +async def test_the_hook_drops_the_slot_only_once_the_origin_changes() -> None: + import httpx + + hook = credential_redirect_hook("https://upstream.example.com/mcp", "esb-oauth") + + same = httpx.Request("GET", "https://upstream.example.com/other", headers={"esb-oauth": "Bearer x"}) + await hook(same) + assert same.headers["esb-oauth"] == "Bearer x" + + foreign = httpx.Request("GET", "https://attacker.example.com/x", headers={"esb-oauth": "Bearer x"}) + await hook(foreign) + assert "esb-oauth" not in foreign.headers diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/IdJagFormFields.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/IdJagFormFields.tsx index df1d8d3436a..70d1bc40c18 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/IdJagFormFields.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/IdJagFormFields.tsx @@ -3,6 +3,7 @@ import React from "react"; import { SimpleTooltip } from "@/components/ui/tooltip"; import { MountedFormField } from "@/components/common_components/MountedFormField"; +import UpstreamTokenHeaderField from "./UpstreamTokenHeaderField"; import { requiredRule } from "@/components/common_components/formRules"; import { MultiSelect } from "@/components/shared/MultiSelect"; import { PasswordInput } from "@/components/shared/PasswordInput"; @@ -205,6 +206,7 @@ const IdJagFormFields: React.FC = ({ isEditing = false }) > {(control) => } + ); }; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.test.tsx index 21bb2801e8c..9316cfa077c 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.test.tsx @@ -286,4 +286,42 @@ describe("OAuthFormFields", () => { }); }); }); + + describe("token header field", () => { + it("renders on the M2M flow", () => { + render( + + + , + ); + expect(screen.getByPlaceholderText("Authorization")).toBeInTheDocument(); + }); + + it("renders on the interactive flow", () => { + render( + + + , + ); + expect(screen.getByPlaceholderText("Authorization")).toBeInTheDocument(); + }); + + it("submits its value under credentials.upstream_token_header", async () => { + const onFinish = vi.fn(); + render( + + + , + ); + fireEvent.change(screen.getByPlaceholderText("Authorization"), { target: { value: "esb-oauth" } }); + fireEvent.click(screen.getByText("Submit")); + await waitFor(() => { + expect(onFinish).toHaveBeenCalledWith( + expect.objectContaining({ + credentials: expect.objectContaining({ upstream_token_header: "esb-oauth" }), + }), + ); + }); + }); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.tsx index 76d667039d4..41817f8c916 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/OAuthFormFields.tsx @@ -11,6 +11,7 @@ import { OAUTH_FLOW } from "@/components/mcp_tools/types"; import { MountedFormField } from "@/components/common_components/MountedFormField"; import { requiredRule } from "@/components/common_components/formRules"; import TokenEndpointAuthMethodField from "./TokenEndpointAuthMethodField"; +import UpstreamTokenHeaderField from "./UpstreamTokenHeaderField"; import { numberControl, parsesAsJson, @@ -175,6 +176,7 @@ const OAuthFormFields: React.FC = ({ {(control) => } + ) : ( <> @@ -237,6 +239,7 @@ const OAuthFormFields: React.FC = ({ {(control) => } + = ({ isEdi /> )} + ); }; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/UpstreamTokenHeaderField.tsx b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/UpstreamTokenHeaderField.tsx new file mode 100644 index 00000000000..154d55c8cb4 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/UpstreamTokenHeaderField.tsx @@ -0,0 +1,31 @@ +import { Info } from "lucide-react"; +import React from "react"; +import { SimpleTooltip } from "@/components/ui/tooltip"; +import { Input } from "@/components/ui/input"; + +import { MountedFormField } from "@/components/common_components/MountedFormField"; +import { textControl } from "./mcpFieldRules"; + +const UpstreamTokenHeaderField: React.FC = () => ( + + Token Header (optional) + + + + + } + name={["credentials", "upstream_token_header"]} + > + {(control) => ( + + )} + +); + +export default UpstreamTokenHeaderField; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.cases.ts b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.cases.ts index c350ac085b6..ef8f728a609 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.cases.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.cases.ts @@ -255,13 +255,28 @@ export const CASES: readonly DifferentialCase[] = [ }, // --- credentials filtering --- - // ADMIN_CONFIG_CREDENTIAL_KEYS is exactly ["upstream_resource"], so only that key - // takes the blank-to-explicit-null branch. A blank client_id is dropped instead. + // Only a key in ADMIN_CONFIG_CREDENTIAL_KEYS takes the blank-to-explicit-null branch, which is + // what makes it clearable: the backend merge preserves an omitted key forever. A blank client_id + // is dropped instead. { label: "blank upstream_resource becomes an explicit null", values: { ...ROOT, auth_type: "oauth2", credentials: { upstream_resource: "", client_secret: "keep" } }, ui: {}, }, + { + label: "blank upstream_token_header becomes an explicit null", + values: { ...ROOT, auth_type: "oauth2", credentials: { upstream_token_header: "", client_secret: "keep" } }, + ui: {}, + }, + { + label: "a set upstream_token_header rides the credentials blob", + values: { + ...ROOT, + auth_type: "oauth2", + credentials: { upstream_token_header: "esb-oauth", client_secret: "keep" }, + }, + ui: {}, + }, { label: "blank non-admin credential is dropped, not nulled", values: { ...ROOT, auth_type: "oauth2", credentials: { client_id: "", client_secret: "keep", scopes: [] } }, diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.test.ts index dd9c8db6d30..13aec81d9e8 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.test.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.test.ts @@ -262,7 +262,14 @@ describe("edit root: exact mounted set per auth configuration", () => { ...PERMS, "delegate_auth_to_upstream", ], - credentials: ["client_id", "client_secret", "token_endpoint_auth_method", "scopes", "upstream_resource"], + credentials: [ + "client_id", + "client_secret", + "token_endpoint_auth_method", + "scopes", + "upstream_resource", + "upstream_token_header", + ], }, ); }); @@ -286,7 +293,14 @@ describe("edit root: exact mounted set per auth configuration", () => { ...PERMS, "delegate_auth_to_upstream", ], - credentials: ["client_id", "client_secret", "scopes", "upstream_resource", "token_endpoint_auth_method"], + credentials: [ + "client_id", + "client_secret", + "scopes", + "upstream_resource", + "token_endpoint_auth_method", + "upstream_token_header", + ], }, ); }); @@ -306,7 +320,7 @@ describe("edit root: exact mounted set per auth configuration", () => { "env_vars", ...PERMS, ], - credentials: ["client_id", "client_secret", "scopes"], + credentials: ["client_id", "client_secret", "scopes", "upstream_token_header"], }, ); }); @@ -324,7 +338,7 @@ describe("edit root: exact mounted set per auth configuration", () => { "env_vars", ...PERMS, ], - credentials: ["client_id", "client_secret", "scopes"], + credentials: ["client_id", "client_secret", "scopes", "upstream_token_header"], }, ); }); @@ -344,6 +358,7 @@ describe("edit root: exact mounted set per auth configuration", () => { ...PERMS, ], credentials: [ + "upstream_token_header", "id_jag_resource_token_endpoint", "client_id", "client_secret", @@ -434,7 +449,14 @@ describe("create root: exact mounted set per configuration", () => { ...PERMS, "delegate_auth_to_upstream", ], - credentials: ["client_id", "client_secret", "scopes", "upstream_resource", "token_endpoint_auth_method"], + credentials: [ + "client_id", + "client_secret", + "scopes", + "upstream_resource", + "token_endpoint_auth_method", + "upstream_token_header", + ], }, ); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.ts b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.ts index 22f0146afc9..af9cbb58b2b 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/mountedServerFields.ts @@ -24,6 +24,7 @@ const OAUTH_M2M_CREDENTIALS = [ "token_endpoint_auth_method", "scopes", "upstream_resource", + "upstream_token_header", ] as const; const OAUTH_INTERACTIVE_CREDENTIALS = [ @@ -32,6 +33,7 @@ const OAUTH_INTERACTIVE_CREDENTIALS = [ "scopes", "upstream_resource", "token_endpoint_auth_method", + "upstream_token_header", ] as const; const OAUTH_INTERACTIVE_ROOT = [ @@ -44,6 +46,7 @@ const OAUTH_INTERACTIVE_ROOT = [ ] as const; const ID_JAG_CREDENTIALS = [ + "upstream_token_header", "id_jag_resource_token_endpoint", "client_id", "client_secret", @@ -100,7 +103,7 @@ const authSubtreeCredentials = ({ authType, oauthFlowType }: AuthSubtreeGates): ]; } if (authType === AUTH_TYPE.OAUTH2_TOKEN_EXCHANGE) { - return [...authValue, "client_id", "client_secret", "scopes"]; + return [...authValue, "client_id", "client_secret", "scopes", "upstream_token_header"]; } if (authType === AUTH_TYPE.OAUTH2_ID_JAG) { return [...authValue, ...ID_JAG_CREDENTIALS]; diff --git a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx index 67e51ab23aa..bc14e7a87ec 100644 --- a/ui/litellm-dashboard/src/components/mcp_tools/types.tsx +++ b/ui/litellm-dashboard/src/components/mcp_tools/types.tsx @@ -151,7 +151,7 @@ const DECLARED_APP_CREDENTIAL_KEYS = ["client_id", "client_secret"] as const; // would destroy admin input), but it must stay OUT of the declared-app set: whether an app exists is // a distinct question that gates the "app may not match upstream" warning, and a server using dynamic // client registration can set a resource indicator while having no app at all. -export const ADMIN_CONFIG_CREDENTIAL_KEYS = ["upstream_resource"] as const; +export const ADMIN_CONFIG_CREDENTIAL_KEYS = ["upstream_resource", "upstream_token_header"] as const; // Minted token material the oauth2 authorize path writes beside the app keys; stripped from restored // snapshots and from any credentials that transit to the temp-session preview so a stale token never From 49170695ce7ec7b3baafe8af70db25da20bac79d Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Thu, 27 Aug 2026 14:50:17 -0700 Subject: [PATCH 43/44] test(e2e): drop the fixture helper docstring The why belongs in the commit message and the PR, not above a one-line helper whose name already says what it returns. --- .../test_chat_completions_regression_e2e.py | 7 ------- 1 file changed, 7 deletions(-) diff --git a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py index 584dbce8089..68c0dfab897 100644 --- a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py +++ b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py @@ -88,13 +88,6 @@ OPENAI_VISION_BACKEND = "openai/gpt-4o" def _cat_image_data_url() -> str: - """The vision image as a data URL, read from a fixture we own. - - An https URL would make every vision run depend on a third-party host staying - up and unthrottled, and a 429 from that host reads as a gateway failure. It also - changes what is under test per provider: litellm downloads the image itself for - bedrock, while openai is handed the link and fetches it from its own servers. A - data URL removes the host and puts both providers on the same bytes.""" return "data:image/jpeg;base64," + base64.b64encode(CAT_IMAGE.read_bytes()).decode() From 1665214bbdbfcbb24471905e56959af116ed20cb Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 27 Aug 2026 15:13:42 -0700 Subject: [PATCH 44/44] feat(together_ai): flag prompt caching on GLM-5.3-Flash like its sibling entries --- litellm/model_prices_and_context_window_backup.json | 1 + model_prices_and_context_window.json | 1 + tests/test_litellm/test_together_ai_model_metadata.py | 1 + 3 files changed, 3 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 82e45337b1f..fe2547282af 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -39065,6 +39065,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 82e45337b1f..fe2547282af 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -39065,6 +39065,7 @@ "source": "https://docs.together.ai/docs/serverless-models", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, diff --git a/tests/test_litellm/test_together_ai_model_metadata.py b/tests/test_litellm/test_together_ai_model_metadata.py index bf35f459bb6..626c4a3d8f2 100644 --- a/tests/test_litellm/test_together_ai_model_metadata.py +++ b/tests/test_litellm/test_together_ai_model_metadata.py @@ -120,6 +120,7 @@ def test_together_glm_53_flash_pricing_and_capabilities(cost_map: CostMap): assert info["max_output_tokens"] == 1048575 assert info["supports_function_calling"] is True assert info["supports_parallel_function_calling"] is True + assert info["supports_prompt_caching"] is True assert info["supports_tool_choice"] is True assert info["supports_response_schema"] is True assert info["supports_vision"] is True