From 107b4ec4db64985de0b3651f401b290ea09e81ed Mon Sep 17 00:00:00 2001 From: yassin Date: Sat, 12 Sep 2026 01:12:39 +0000 Subject: [PATCH 001/119] fix(redis): log a timeout streak once per interval instead of one line per cache call Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 140 ++++++++++++------ litellm/constants.py | 3 + tests/test_litellm/caching/test_dual_cache.py | 57 +++++++ .../test_litellm/caching/test_redis_cache.py | 57 +++++++ 4 files changed, 211 insertions(+), 46 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index 2c36995c4f8..eaac7ef7b0b 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -15,6 +15,7 @@ import hashlib import inspect import json import logging +import threading import time from collections.abc import Awaitable, Callable, Sequence from contextvars import ContextVar @@ -32,6 +33,7 @@ from litellm.constants import ( REDIS_CIRCUIT_BREAKER_FAILURE_THRESHOLD, REDIS_CIRCUIT_BREAKER_RECOVERY_TIMEOUT, REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION, + REDIS_TIMEOUT_LOG_INTERVAL, ) from litellm.litellm_core_utils.core_helpers import _get_parent_otel_span_from_kwargs from litellm.litellm_core_utils.coroutine_checker import coroutine_checker @@ -404,13 +406,58 @@ class RedisCircuitBreakerOpenError(Exception): pass +class _RedisTimeoutLogThrottle: + """Admits one Redis timeout log line per interval and counts the timeouts it suppressed in between.""" + + def __init__(self, interval: float, clock: Callable[[], float] = time.time) -> None: + self.interval = interval + self._clock = clock + self._lock = threading.Lock() + self._last_logged_at: float | None = None + self._suppressed = 0 + + def admit(self) -> int | None: + """Return the number of timeouts suppressed since the last admitted line, or None to suppress this one.""" + with self._lock: + now: Final = self._clock() + if self._last_logged_at is not None and now - self._last_logged_at < self.interval: + self._suppressed += 1 + return None + suppressed: Final = self._suppressed + self._suppressed = 0 + self._last_logged_at = now + return suppressed + + +_redis_timeout_log_throttle: Final = _RedisTimeoutLogThrottle(REDIS_TIMEOUT_LOG_INTERVAL) + + def log_redis_failure( logger: logging.Logger, level: int, message: str, exc: BaseException, with_traceback: bool = False ) -> None: if isinstance(exc, RedisCircuitBreakerOpenError): - logger.debug("%s: %s", message, exc) + logger.debug("%s: %s", message, exc, stacklevel=2) return - logger.log(level, "%s: %s", message, exc, exc_info=exc if with_traceback else None) + exc_info: Final = exc if with_traceback else None + if not _is_redis_timeout_failure(exc): + logger.log(level, "%s: %s", message, exc, exc_info=exc_info, stacklevel=2) + return + suppressed: Final = _redis_timeout_log_throttle.admit() + if suppressed is None: + logger.debug("%s: %s", message, exc, stacklevel=2) + return + if suppressed == 0: + logger.log(level, "%s: %s", message, exc, exc_info=exc_info, stacklevel=2) + return + logger.log( + level, + "%s: %s (%d more Redis timeouts since the previous Redis timeout line were logged at DEBUG)", + message, + exc, + suppressed, + exc_info=exc_info, + stacklevel=2, + ) @dataclass(frozen=True, slots=True) @@ -783,10 +830,8 @@ class RedisCache(BaseCache): ## LOGGING ## end_time = time.time() _duration = end_time - start_time - verbose_logger.error( - "LiteLLM Redis Caching: increment_cache() - Got exception from REDIS %s, Writing value=%s", - str(e), - value, + log_redis_failure( + verbose_logger, logging.ERROR, "LiteLLM Redis Caching: increment_cache() - Got exception from REDIS", e ) raise e @@ -992,11 +1037,8 @@ class RedisCache(BaseCache): call_type=f"async_set_cache <- {_get_call_stack_info()}", ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, key=%r, value=%r", - str(e), - key, - value, + log_redis_failure( + verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e ) raise e @@ -1044,10 +1086,8 @@ class RedisCache(BaseCache): event_metadata={"key": key}, ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, Writing value=%s", - str(e), - value, + log_redis_failure( + verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e ) _record_swallowed_redis_failure(self._circuit_breaker, e) @@ -1094,7 +1134,6 @@ class RedisCache(BaseCache): start_time: Final = time.time() print_verbose(f"Set Async Redis Cache: key list: {cache_list}\nttl={ttl}, redis_version={self.redis_version}") - cache_value: Final = None try: async with _redis_client.pipeline(transaction=False) as pipe: results: Final = await self._pipeline_helper(pipe, cache_list, ttl) @@ -1131,10 +1170,11 @@ class RedisCache(BaseCache): ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async set_cache_pipeline() - Got exception from REDIS %s, Writing value=%s", - str(e), - cache_value, + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async set_cache_pipeline() - Got exception from REDIS", + e, ) _record_swallowed_redis_failure(self._circuit_breaker, e) @@ -1177,10 +1217,8 @@ class RedisCache(BaseCache): ) ) # NON blocking - notify users Redis is throwing an exception - verbose_logger.error( - "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, Writing value=%s", - str(e), - value, + log_redis_failure( + verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e ) raise e @@ -1216,10 +1254,11 @@ class RedisCache(BaseCache): ) ) # NON blocking - notify users Redis is throwing an exception - verbose_logger.error( - "LiteLLM Redis Caching: async set_cache_sadd() - Got exception from REDIS %s, Writing value=%s", - str(e), - value, + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async set_cache_sadd() - Got exception from REDIS", + e, ) _record_swallowed_redis_failure(self._circuit_breaker, e) @@ -1288,10 +1327,11 @@ class RedisCache(BaseCache): parent_otel_span=parent_otel_span, ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async async_increment() - Got exception from REDIS %s, Writing value=%s", - str(e), - value, + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async async_increment() - Got exception from REDIS", + e, ) raise e @@ -1377,7 +1417,9 @@ class RedisCache(BaseCache): print_verbose(f"Got Redis Cache: key: {key}, cached_response {cached_response}") return self._get_cache_logic(cached_response=cached_response) except Exception as e: - verbose_logger.error("litellm.caching.caching: get() - Got exception from REDIS: %s", e) + log_redis_failure( + verbose_logger, logging.ERROR, "litellm.caching.caching: get() - Got exception from REDIS", e + ) _record_swallowed_redis_failure(self._circuit_breaker, e) def _run_redis_mget_operation(self, keys: list[str]) -> Sequence[bytes | str | None]: @@ -1455,7 +1497,7 @@ class RedisCache(BaseCache): end_time=failed_at, parent_otel_span=parent_otel_span, ) - verbose_logger.error("Error occurred in batch get cache - %s", e) + log_redis_failure(verbose_logger, logging.ERROR, "Error occurred in batch get cache", e) _record_swallowed_redis_failure(self._circuit_breaker, e) return key_value_dict @@ -1574,7 +1616,7 @@ class RedisCache(BaseCache): parent_otel_span=parent_otel_span, ) ) - verbose_logger.error("Error occurred in async batch get cache - %s", e) + log_redis_failure(verbose_logger, logging.ERROR, "Error occurred in async batch get cache", e) _record_swallowed_redis_failure(self._circuit_breaker, e) return key_value_dict @@ -1799,9 +1841,11 @@ class RedisCache(BaseCache): parent_otel_span=_get_parent_otel_span_from_kwargs(kwargs), ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async increment_pipeline() - Got exception from REDIS %s", - str(e), + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async increment_pipeline() - Got exception from REDIS", + e, ) raise e @@ -1878,7 +1922,7 @@ class RedisCache(BaseCache): call_type=f"async_rpush <- {_get_call_stack_info()}", ) ) - verbose_logger.error("LiteLLM Redis Cache RPUSH: - Got exception from REDIS : %s", e) + log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Redis Cache RPUSH: - Got exception from REDIS", e) raise e async def _pipeline_rpush_helper( @@ -1946,9 +1990,11 @@ class RedisCache(BaseCache): call_type=f"async_rpush_pipeline <- {_get_call_stack_info()}", ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async_rpush_pipeline() - Got exception from REDIS %s", - str(e), + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async_rpush_pipeline() - Got exception from REDIS", + e, ) raise e @@ -2024,7 +2070,7 @@ class RedisCache(BaseCache): call_type=f"async_lpop <- {_get_call_stack_info()}", ) ) - verbose_logger.error("LiteLLM Redis Cache LPOP: - Got exception from REDIS : %s", e) + log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Redis Cache LPOP: - Got exception from REDIS", e) raise e async def _pipeline_lpop_helper( @@ -2135,8 +2181,10 @@ class RedisCache(BaseCache): call_type=f"async_lpop_pipeline <- {_get_call_stack_info()}", ) ) - verbose_logger.error( - "LiteLLM Redis Caching: async_lpop_pipeline() - Got exception from REDIS %s", - str(e), + log_redis_failure( + verbose_logger, + logging.ERROR, + "LiteLLM Redis Caching: async_lpop_pipeline() - Got exception from REDIS", + e, ) raise e diff --git a/litellm/constants.py b/litellm/constants.py index 6b984c2673c..a32551b4480 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -459,6 +459,9 @@ REDIS_CIRCUIT_BREAKER_ENABLED: Final = os.getenv("REDIS_CIRCUIT_BREAKER_ENABLED" # minimum seconds a timeout-only failure streak must span before it can open the breaker, # so one event-loop stall timing out many queued calls at once does not trip it REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION: Final = float(os.getenv("REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION", 5.0)) +# seconds between Redis timeout log lines: the first timeout of a streak logs at the caller's level, +# later ones log at DEBUG until the interval passes and one line summarizes how many were suppressed +REDIS_TIMEOUT_LOG_INTERVAL: Final = float(os.getenv("REDIS_TIMEOUT_LOG_INTERVAL", "5.0")) # Seconds of idle before a Redis cluster connection is validated with a PING and # reconnected if dead, so a connection silently dropped by a cluster restart # (e.g. ElastiCache Serverless maintenance) is not reused while broken diff --git a/tests/test_litellm/caching/test_dual_cache.py b/tests/test_litellm/caching/test_dual_cache.py index 4c9068722b8..850fa14106b 100644 --- a/tests/test_litellm/caching/test_dual_cache.py +++ b/tests/test_litellm/caching/test_dual_cache.py @@ -704,3 +704,60 @@ async def test_open_breaker_keeps_async_batch_read_memory_hits_and_releases_rese assert list(await cache.async_batch_get_cache(["k1", "k2"])) == ["v1", None] assert "k2" not in cache.last_redis_batch_access_time + + +@pytest.mark.asyncio +async def test_redis_timeouts_falling_back_to_memory_log_once_per_interval(caplog, monkeypatch): + """The in-memory fallback WARNING must not repeat for every timed-out increment during a blip. + + The rate limiter's pipeline increments and the dual cache increments each logged a WARNING per + call while Redis timed out, hundreds of lines per second before the breaker opened. The first + timeout of a streak keeps its WARNING, the rest are DEBUG until the summary interval passes. + """ + from redis.exceptions import TimeoutError as RedisTimeoutError + + from litellm.caching import redis_cache as redis_cache_module + from litellm.caching.redis_cache import _RedisTimeoutLogThrottle + + clock = MagicMock(return_value=1_000.0) + monkeypatch.setattr( + redis_cache_module, "_redis_timeout_log_throttle", _RedisTimeoutLogThrottle(interval=5.0, clock=clock) + ) + + class _TimingOutRedis: + async def async_increment_pipeline(self, increment_list, **kwargs): + raise RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + + async def async_increment(self, key, value, **kwargs): + raise RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + + cache = DualCache(in_memory_cache=InMemoryCache(), redis_cache=_TimingOutRedis()) # pyright: ignore[reportArgumentType] # duck-typed Redis double + increments = [RedisPipelineIncrementOperation(key="k", increment_value=1.0, ttl=60)] + + with caplog.at_level(logging.DEBUG, logger="LiteLLM"): + for _ in range(100): + await cache.async_increment_cache_pipeline(increment_list=increments) + await cache.async_increment_cache("k", 1.0) + + visible = [r for r in caplog.records if r.levelno >= logging.WARNING] + assert [(r.levelno, r.getMessage()) for r in visible] == [ + ( + logging.WARNING, + "Redis async_increment_cache_pipeline failed, falling back to in-memory result:" + " Timeout reading from 127.0.0.1:6379", + ) + ] + assert visible[0].filename == "dual_cache.py" + assert sum("Timeout reading from" in r.getMessage() for r in caplog.records) == 200 + + caplog.clear() + clock.return_value += 5.0 + with caplog.at_level(logging.DEBUG, logger="LiteLLM"): + await cache.async_increment_cache("k", 1.0) + assert [(r.levelno, r.getMessage()) for r in caplog.records] == [ + ( + logging.WARNING, + "Redis async_increment_cache failed, falling back to in-memory result: Timeout reading from 127.0.0.1:6379" + " (199 more Redis timeouts since the previous Redis timeout line were logged at DEBUG)", + ) + ] diff --git a/tests/test_litellm/caching/test_redis_cache.py b/tests/test_litellm/caching/test_redis_cache.py index bcae33b976e..d0974b2420c 100644 --- a/tests/test_litellm/caching/test_redis_cache.py +++ b/tests/test_litellm/caching/test_redis_cache.py @@ -1202,3 +1202,60 @@ async def test_a_probe_overtaken_by_a_later_outage_leaves_the_breaker_to_the_new new_probe_release.set() assert await new_probe == "new probe" assert breaker._state == breaker.CLOSED + + +def test_timeouts_during_a_blip_log_once_per_interval_not_once_per_call(sync_batch_redis_cache, caplog, monkeypatch): + """A Redis latency blip must not write one ERROR line per timed-out cache call. + + Before the breaker opens (up to REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION of timeouts) every + cache operation logged its own ERROR or WARNING line, so one single-worker proxy wrote + ~1100 lines in 5 s at LITELLM_LOG=WARNING. A timeout streak now logs its first failure, then + one summary line per REDIS_TIMEOUT_LOG_INTERVAL carrying the count of suppressed timeouts, + while every timeout stays visible at DEBUG. Hard connectivity failures keep their per-call line. + """ + import logging + + from redis.exceptions import TimeoutError as RedisTimeoutError + + from litellm.caching import redis_cache as redis_cache_module + from litellm.caching.redis_cache import _RedisTimeoutLogThrottle + + clock = MagicMock(return_value=1_000.0) + monkeypatch.setattr( + redis_cache_module, "_redis_timeout_log_throttle", _RedisTimeoutLogThrottle(interval=5.0, clock=clock) + ) + sync_batch_redis_cache.redis_client.get.side_effect = RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + sync_batch_redis_cache.redis_client.mget.side_effect = RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + + with caplog.at_level(logging.DEBUG, logger="LiteLLM"): + for _ in range(200): + assert sync_batch_redis_cache.get_cache("lit7520") is None + assert sync_batch_redis_cache.batch_get_cache(key_list=["lit7520"]) == {} + + timeout_records = [r for r in caplog.records if "Timeout reading from" in r.getMessage()] + assert len(timeout_records) == 201, "every timeout must stay visible at DEBUG" + assert [r.getMessage() for r in timeout_records if r.levelno >= logging.WARNING] == [ + "litellm.caching.caching: get() - Got exception from REDIS: Timeout reading from 127.0.0.1:6379" + ] + assert timeout_records[0].levelno == logging.ERROR + assert timeout_records[0].filename == "redis_cache.py" + assert timeout_records[0].lineno != timeout_records[-1].lineno, "the record must point at the cache operation" + + caplog.clear() + clock.return_value += 5.0 + with caplog.at_level(logging.DEBUG, logger="LiteLLM"): + assert sync_batch_redis_cache.batch_get_cache(key_list=["lit7520"]) == {} + assert [(r.levelno, r.getMessage()) for r in caplog.records] == [ + ( + logging.ERROR, + "Error occurred in batch get cache: Timeout reading from 127.0.0.1:6379" + " (200 more Redis timeouts since the previous Redis timeout line were logged at DEBUG)", + ) + ] + + caplog.clear() + sync_batch_redis_cache.redis_client.get.side_effect = OSError("redis unavailable") + with caplog.at_level(logging.DEBUG, logger="LiteLLM"): + for _ in range(3): + assert sync_batch_redis_cache.get_cache("lit7520") is None + assert [r.levelno for r in caplog.records if "redis unavailable" in r.getMessage()] == [logging.ERROR] * 3 From 9c84e98fb22bd0f6e2c359f335bbc329181bb8bd Mon Sep 17 00:00:00 2001 From: yassin Date: Sat, 12 Sep 2026 01:25:32 +0000 Subject: [PATCH 002/119] fix(proxy): treat a Redis timeout in spend counter increments as an already-logged cache failure The cost tracking callback logged its own ERROR with a traceback for every request whose spend counter increment timed out, on top of the cache layer's throttled line. Timeouts now take the same path as breaker-open refusals: invalidate the counters and return. Also exposes is_redis_timeout_failure publicly for that caller and drops the comment on the new constant Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 10 +++---- litellm/constants.py | 2 -- litellm/proxy/proxy_server.py | 4 +-- .../test_litellm/caching/test_redis_cache.py | 26 +++++++++---------- .../proxy/proxy_server/test_spend_counters.py | 22 ++++++++++++++++ 5 files changed, 42 insertions(+), 22 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index eaac7ef7b0b..7a3e689a667 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -330,7 +330,7 @@ def _redis_timeout_error_types() -> tuple[type, ...]: return (RedisTimeoutError, TimeoutError) -def _is_redis_timeout_failure(exc: BaseException) -> bool: +def is_redis_timeout_failure(exc: BaseException) -> bool: return isinstance(exc, _redis_timeout_error_types()) @@ -398,7 +398,7 @@ def _record_swallowed_redis_failure(breaker: RedisCircuitBreaker, exc: BaseExcep """ if not _is_redis_health_failure(exc): return - breaker.record_failure(is_timeout=_is_redis_timeout_failure(exc)) + breaker.record_failure(is_timeout=is_redis_timeout_failure(exc)) _swallowed_redis_failures.set(_swallowed_redis_failures.get() + 1) @@ -439,7 +439,7 @@ def log_redis_failure( logger.debug("%s: %s", message, exc, stacklevel=2) return exc_info: Final = exc if with_traceback else None - if not _is_redis_timeout_failure(exc): + if not is_redis_timeout_failure(exc): logger.log(level, "%s: %s", message, exc, exc_info=exc_info, stacklevel=2) return suppressed: Final = _redis_timeout_log_throttle.admit() @@ -504,7 +504,7 @@ async def _run_under_circuit_breaker( result: Final = await call() except Exception as e: if _is_redis_health_failure(e): - breaker.record_failure(is_timeout=_is_redis_timeout_failure(e)) + breaker.record_failure(is_timeout=is_redis_timeout_failure(e)) raise _exit_circuit_breaker(breaker, admission) return result @@ -521,7 +521,7 @@ def _run_under_circuit_breaker_sync( result: Final = call() except Exception as e: if _is_redis_health_failure(e): - breaker.record_failure(is_timeout=_is_redis_timeout_failure(e)) + breaker.record_failure(is_timeout=is_redis_timeout_failure(e)) raise _exit_circuit_breaker(breaker, admission) return result diff --git a/litellm/constants.py b/litellm/constants.py index a32551b4480..60e1c682238 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -459,8 +459,6 @@ REDIS_CIRCUIT_BREAKER_ENABLED: Final = os.getenv("REDIS_CIRCUIT_BREAKER_ENABLED" # minimum seconds a timeout-only failure streak must span before it can open the breaker, # so one event-loop stall timing out many queued calls at once does not trip it REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION: Final = float(os.getenv("REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION", 5.0)) -# seconds between Redis timeout log lines: the first timeout of a streak logs at the caller's level, -# later ones log at DEBUG until the interval passes and one line summarizes how many were suppressed REDIS_TIMEOUT_LOG_INTERVAL: Final = float(os.getenv("REDIS_TIMEOUT_LOG_INTERVAL", "5.0")) # Seconds of idle before a Redis cluster connection is validated with a PING and # reconnected if dead, so a connection silently dropped by a cluster restart diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index c8fd7edfed6..318ea96dbcb 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -251,7 +251,7 @@ import litellm._redis from litellm import Router from litellm._logging import _redact_string, verbose_proxy_logger, verbose_router_logger from litellm.caching.caching import DualCache, RedisCache -from litellm.caching.redis_cache import RedisCircuitBreakerOpenError +from litellm.caching.redis_cache import RedisCircuitBreakerOpenError, is_redis_timeout_failure from litellm.caching.redis_cluster_cache import RedisClusterCache from litellm.constants import ( _REALTIME_BODY_CACHE_SIZE, @@ -3411,7 +3411,7 @@ async def _apply_spend_counter_increments(pending: Sequence[_PendingSpendIncreme results: Final = await redis_cache.async_increment_pipeline(increment_list=increment_list) except Exception as e: await asyncio.gather(*(_invalidate_spend_counter(counter_key=item.counter_key) for item in pending)) - if isinstance(e, RedisCircuitBreakerOpenError): + if isinstance(e, RedisCircuitBreakerOpenError) or is_redis_timeout_failure(e): return raise for item, current_value in zip(pending, results or ()): diff --git a/tests/test_litellm/caching/test_redis_cache.py b/tests/test_litellm/caching/test_redis_cache.py index d0974b2420c..5840f450ac6 100644 --- a/tests/test_litellm/caching/test_redis_cache.py +++ b/tests/test_litellm/caching/test_redis_cache.py @@ -977,17 +977,17 @@ async def test_stale_timeout_does_not_let_sub_threshold_hard_failures_open_the_b from redis.exceptions import ConnectionError as RedisConnectionError from redis.exceptions import TimeoutError as RedisTimeoutError - from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure + from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure breaker = RedisCircuitBreaker(failure_threshold=3, recovery_timeout=60, timeout_min_duration=0.05) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out"))) await asyncio.sleep(0.06) for _ in range(breaker.failure_threshold - 1): - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused"))) assert breaker.is_open() is False, "2 hard failures and 1 stale timeout are below both thresholds" - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused"))) assert breaker.is_open() is True, "the threshold-th hard failure must still open it" @@ -999,19 +999,19 @@ async def test_hard_failure_resets_timeout_streak_so_a_later_burst_must_earn_its from redis.exceptions import ConnectionError as RedisConnectionError from redis.exceptions import TimeoutError as RedisTimeoutError - from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure + from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure breaker = RedisCircuitBreaker(failure_threshold=3, recovery_timeout=60, timeout_min_duration=0.05) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out"))) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused"))) await asyncio.sleep(0.06) for _ in range(breaker.failure_threshold): - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out"))) assert breaker.is_open() is False, "the burst is instantaneous, so the duration gate must hold it closed" await asyncio.sleep(0.06) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out"))) assert breaker.is_open() is True, "the same run of timeouts persisting past the duration must open it" @@ -1022,7 +1022,7 @@ async def test_breaker_metrics_track_state_and_failure_class(): from redis.exceptions import ConnectionError as RedisConnectionError from redis.exceptions import TimeoutError as RedisTimeoutError - from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure + from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure def sample(name, labels=None): return REGISTRY.get_sample_value(name, labels) or 0.0 @@ -1034,9 +1034,9 @@ async def test_breaker_metrics_track_state_and_failure_class(): closed_gauge_before = sample("litellm_redis_circuit_breaker_state", {"state": "closed"}) breaker = RedisCircuitBreaker(failure_threshold=2, recovery_timeout=60, timeout_min_duration=5.0) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("t"))) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused"))) - breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("t"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused"))) + breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused"))) assert sample("litellm_redis_circuit_breaker_failures_total", {"failure_class": "timeout"}) == timeout_before + 1 assert sample("litellm_redis_circuit_breaker_failures_total", {"failure_class": "connectivity"}) == hard_before + 2 diff --git a/tests/test_litellm/proxy/proxy_server/test_spend_counters.py b/tests/test_litellm/proxy/proxy_server/test_spend_counters.py index 2f47736a398..19b5a11af33 100644 --- a/tests/test_litellm/proxy/proxy_server/test_spend_counters.py +++ b/tests/test_litellm/proxy/proxy_server/test_spend_counters.py @@ -1180,6 +1180,28 @@ async def test_apply_spend_counter_increments_open_breaker_invalidates_and_retur fake_cache.in_memory_cache.set_cache.assert_not_called() +@pytest.mark.asyncio +async def test_apply_spend_counter_increments_redis_timeout_invalidates_and_returns(monkeypatch): + """A Redis timeout is the streak the breaker is already counting and the cache layer already logged. + + Re-raising it sent every request in the pre-open window through the cost callback's error + path, which logged a traceback and fired the failed-tracking alert once per request. + """ + from redis.exceptions import TimeoutError as RedisTimeoutError + + fake_cache = _make_spend_counter_cache() + fake_cache.redis_cache.async_increment_pipeline = AsyncMock( + side_effect=RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + ) + monkeypatch.setattr(ps, "spend_counter_cache", fake_cache) + + await ps._apply_spend_counter_increments(_two_pending_increments()) + + deleted_keys = sorted(call.kwargs["key"] for call in fake_cache.in_memory_cache.delete_cache.call_args_list) + assert deleted_keys == ["spend:key:k", "spend:team:t"] + fake_cache.in_memory_cache.set_cache.assert_not_called() + + @pytest.mark.asyncio async def test_apply_spend_counter_increments_other_redis_error_invalidates_and_raises(monkeypatch): fake_cache = _make_spend_counter_cache() From f681a978f06baa13da0f0c24f7b1ac3a20d9a02a Mon Sep 17 00:00:00 2001 From: yassin Date: Sat, 12 Sep 2026 01:37:36 +0000 Subject: [PATCH 003/119] fix(redis): use a monotonic clock for the timeout log throttle and trim test docstrings Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/caching/redis_cache.py | 2 +- tests/test_litellm/caching/test_dual_cache.py | 7 +------ tests/test_litellm/caching/test_redis_cache.py | 9 +-------- .../proxy/proxy_server/test_spend_counters.py | 6 +----- 4 files changed, 4 insertions(+), 20 deletions(-) diff --git a/litellm/caching/redis_cache.py b/litellm/caching/redis_cache.py index 7a3e689a667..e5e27d1e02b 100644 --- a/litellm/caching/redis_cache.py +++ b/litellm/caching/redis_cache.py @@ -409,7 +409,7 @@ class RedisCircuitBreakerOpenError(Exception): class _RedisTimeoutLogThrottle: """Admits one Redis timeout log line per interval and counts the timeouts it suppressed in between.""" - def __init__(self, interval: float, clock: Callable[[], float] = time.time) -> None: + def __init__(self, interval: float, clock: Callable[[], float] = time.monotonic) -> None: self.interval = interval self._clock = clock self._lock = threading.Lock() diff --git a/tests/test_litellm/caching/test_dual_cache.py b/tests/test_litellm/caching/test_dual_cache.py index 850fa14106b..6f29be00b30 100644 --- a/tests/test_litellm/caching/test_dual_cache.py +++ b/tests/test_litellm/caching/test_dual_cache.py @@ -708,12 +708,7 @@ async def test_open_breaker_keeps_async_batch_read_memory_hits_and_releases_rese @pytest.mark.asyncio async def test_redis_timeouts_falling_back_to_memory_log_once_per_interval(caplog, monkeypatch): - """The in-memory fallback WARNING must not repeat for every timed-out increment during a blip. - - The rate limiter's pipeline increments and the dual cache increments each logged a WARNING per - call while Redis timed out, hundreds of lines per second before the breaker opened. The first - timeout of a streak keeps its WARNING, the rest are DEBUG until the summary interval passes. - """ + """The first fallback WARNING of a timeout streak logs, the rest stay at DEBUG until the summary.""" from redis.exceptions import TimeoutError as RedisTimeoutError from litellm.caching import redis_cache as redis_cache_module diff --git a/tests/test_litellm/caching/test_redis_cache.py b/tests/test_litellm/caching/test_redis_cache.py index 5840f450ac6..4ca33894aed 100644 --- a/tests/test_litellm/caching/test_redis_cache.py +++ b/tests/test_litellm/caching/test_redis_cache.py @@ -1205,14 +1205,7 @@ async def test_a_probe_overtaken_by_a_later_outage_leaves_the_breaker_to_the_new def test_timeouts_during_a_blip_log_once_per_interval_not_once_per_call(sync_batch_redis_cache, caplog, monkeypatch): - """A Redis latency blip must not write one ERROR line per timed-out cache call. - - Before the breaker opens (up to REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION of timeouts) every - cache operation logged its own ERROR or WARNING line, so one single-worker proxy wrote - ~1100 lines in 5 s at LITELLM_LOG=WARNING. A timeout streak now logs its first failure, then - one summary line per REDIS_TIMEOUT_LOG_INTERVAL carrying the count of suppressed timeouts, - while every timeout stays visible at DEBUG. Hard connectivity failures keep their per-call line. - """ + """A timeout streak logs its first failure plus one summary per interval; other failures log per call.""" import logging from redis.exceptions import TimeoutError as RedisTimeoutError diff --git a/tests/test_litellm/proxy/proxy_server/test_spend_counters.py b/tests/test_litellm/proxy/proxy_server/test_spend_counters.py index 19b5a11af33..4a1fc389d3e 100644 --- a/tests/test_litellm/proxy/proxy_server/test_spend_counters.py +++ b/tests/test_litellm/proxy/proxy_server/test_spend_counters.py @@ -1182,11 +1182,7 @@ async def test_apply_spend_counter_increments_open_breaker_invalidates_and_retur @pytest.mark.asyncio async def test_apply_spend_counter_increments_redis_timeout_invalidates_and_returns(monkeypatch): - """A Redis timeout is the streak the breaker is already counting and the cache layer already logged. - - Re-raising it sent every request in the pre-open window through the cost callback's error - path, which logged a traceback and fired the failed-tracking alert once per request. - """ + """A Redis timeout invalidates the counters and returns without reaching the cost callback's error path.""" from redis.exceptions import TimeoutError as RedisTimeoutError fake_cache = _make_spend_counter_cache() From 28f2d1f0168aa31639a23447d391516129267069 Mon Sep 17 00:00:00 2001 From: yassin Date: Sat, 12 Sep 2026 01:58:36 +0000 Subject: [PATCH 004/119] test(redis): cover the write and list timeout paths going through the shared log throttle Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../test_litellm/caching/test_redis_cache.py | 57 +++++++++++++++++++ 1 file changed, 57 insertions(+) diff --git a/tests/test_litellm/caching/test_redis_cache.py b/tests/test_litellm/caching/test_redis_cache.py index 4ca33894aed..5ea21dae539 100644 --- a/tests/test_litellm/caching/test_redis_cache.py +++ b/tests/test_litellm/caching/test_redis_cache.py @@ -1252,3 +1252,60 @@ def test_timeouts_during_a_blip_log_once_per_interval_not_once_per_call(sync_bat for _ in range(3): assert sync_batch_redis_cache.get_cache("lit7520") is None assert [r.levelno for r in caplog.records if "redis unavailable" in r.getMessage()] == [logging.ERROR] * 3 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "call_method", + [ + pytest.param(lambda c: c.async_set_cache_pipeline([("lit7520", "v")]), id="async_set_cache_pipeline"), + pytest.param(lambda c: c.async_set_cache_sadd("lit7520", ["v"], ttl=None), id="async_set_cache_sadd"), + pytest.param(lambda c: c.async_increment("lit7520", 1.0), id="async_increment"), + pytest.param( + lambda c: c.async_increment_pipeline([{"key": "lit7520", "increment_value": 1.0, "ttl": 60}]), + id="async_increment_pipeline", + ), + pytest.param(lambda c: c.async_rpush("lit7520", ["v"]), id="async_rpush"), + pytest.param( + lambda c: c.async_rpush_pipeline([{"key": "lit7520", "values": ["v"]}]), id="async_rpush_pipeline" + ), + pytest.param(lambda c: c.async_lpop("lit7520"), id="async_lpop"), + pytest.param(lambda c: c.async_lpop_pipeline([{"key": "lit7520", "count": 1}]), id="async_lpop_pipeline"), + ], +) +async def test_write_path_timeouts_inside_the_interval_stay_at_debug(call_method, caplog, monkeypatch, redis_no_ping): + """A write or list operation timing out mid-streak is counted by the throttle instead of logging its own ERROR.""" + import contextlib + import logging + + from redis.exceptions import TimeoutError as RedisTimeoutError + + from litellm.caching import redis_cache as redis_cache_module + from litellm.caching.redis_cache import _RedisTimeoutLogThrottle + + clock = MagicMock(return_value=1_000.0) + throttle = _RedisTimeoutLogThrottle(interval=5.0, clock=clock) + assert throttle.admit() == 0 + monkeypatch.setattr(redis_cache_module, "_redis_timeout_log_throttle", throttle) + + timeout = RedisTimeoutError("Timeout reading from 127.0.0.1:6379") + client = MagicMock() + client.pipeline.return_value.__aenter__.side_effect = timeout + client.sadd = AsyncMock(side_effect=timeout) + client.incrbyfloat = AsyncMock(side_effect=timeout) + client.rpush = AsyncMock(side_effect=timeout) + client.lpop = AsyncMock(side_effect=timeout) + monkeypatch.setenv("REDIS_HOST", "https://my-test-host") + cache = RedisCache() + + with ( + patch.object(cache, "init_async_client", return_value=client), + caplog.at_level(logging.DEBUG, logger="LiteLLM"), + ): + with contextlib.suppress(RedisTimeoutError): + await call_method(cache) + + timeout_records = [r for r in caplog.records if "Timeout reading from" in r.getMessage()] + assert [(r.levelno, r.filename) for r in timeout_records] == [(logging.DEBUG, "redis_cache.py")] + clock.return_value += 5.0 + assert throttle.admit() == 1 From cba843cc167bbadf5d18bee0ea07a443a4838ba1 Mon Sep 17 00:00:00 2001 From: Tin Chi Lo Date: Sat, 12 Sep 2026 12:03:04 -0700 Subject: [PATCH 005/119] feat(proxy): predict prompt-cache costs across deployments --- .../llms/anthropic/prompt_cache_prediction.py | 388 ++++++++++ litellm/proxy/_types.py | 1 + litellm/proxy/auth/auth_checks.py | 5 +- litellm/proxy/auth/auth_utils.py | 35 +- litellm/proxy/auth/user_api_key_auth.py | 15 +- .../common_utils/prompt_cache_pricing.py | 91 +++ litellm/proxy/hooks/__init__.py | 2 + .../hooks/parallel_request_limiter_v3.py | 204 ++--- .../proxy/hooks/prompt_cache_prediction.py | 142 ++++ .../cost_tracking_settings.py | 2 + .../prompt_cache_prediction.py | 278 +++++++ .../streaming_handler.py | 18 + .../prompt_cache_prediction.py | 67 ++ .../test_anthropic_prompt_cache_prediction.py | 209 ++++++ .../proxy/auth/test_auth_utils.py | 226 ++++++ .../common_utils/test_prompt_cache_pricing.py | 105 +++ .../hooks/test_parallel_request_limiter_v3.py | 245 ++++++ .../proxy/hooks/test_prompt_cache_observer.py | 300 ++++++++ .../test_prompt_cache_prediction.py | 698 ++++++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 164 ++++ 20 files changed, 3106 insertions(+), 89 deletions(-) create mode 100644 litellm/llms/anthropic/prompt_cache_prediction.py create mode 100644 litellm/proxy/common_utils/prompt_cache_pricing.py create mode 100644 litellm/proxy/hooks/prompt_cache_prediction.py create mode 100644 litellm/proxy/management_endpoints/prompt_cache_prediction.py create mode 100644 litellm/types/management_endpoints/prompt_cache_prediction.py create mode 100644 tests/test_litellm/llms/anthropic/test_anthropic_prompt_cache_prediction.py create mode 100644 tests/test_litellm/proxy/common_utils/test_prompt_cache_pricing.py create mode 100644 tests/test_litellm/proxy/hooks/test_prompt_cache_observer.py create mode 100644 tests/test_litellm/proxy/management_endpoints/test_prompt_cache_prediction.py diff --git a/litellm/llms/anthropic/prompt_cache_prediction.py b/litellm/llms/anthropic/prompt_cache_prediction.py new file mode 100644 index 00000000000..e69a02bd93a --- /dev/null +++ b/litellm/llms/anthropic/prompt_cache_prediction.py @@ -0,0 +1,388 @@ +from __future__ import annotations + +import hashlib +import json +from collections.abc import Mapping, Sequence +from dataclasses import dataclass +from itertools import accumulate +from types import MappingProxyType +from typing import Annotated, Final, Literal, Protocol, TypeAlias + +import httpx +from pydantic import BaseModel, ConfigDict, Field, JsonValue, StrictInt, TypeAdapter, ValidationError + +import litellm +from litellm.llms.anthropic.common_utils import AnthropicModelInfo, is_anthropic_oauth_key +from litellm.llms.anthropic.count_tokens.handler import AnthropicCountTokensHandler +from litellm.llms.anthropic.experimental_pass_through.messages.transformation import DEFAULT_ANTHROPIC_API_VERSION +from litellm.types.router import LiteLLM_Params +from litellm.types.utils import ModelResponse + +_JSON_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) +_HEADERS: Final = TypeAdapter(dict[str, str]) +_counter: Final = AnthropicCountTokensHandler() + + +_NATIVE_HEADERS: Final = frozenset( + ( + "host", + "accept", + "accept-encoding", + "connection", + "user-agent", + "content-length", + "content-type", + "x-api-key", + "anthropic-version", + ) +) + +_DEPLOYMENT_OPTIONS: Final = frozenset( + { + "model", + "api_key", + "api_base", + "custom_llm_provider", + "rpm", + "tpm", + "timeout", + "stream_timeout", + "max_retries", + "num_retries", + "max_parallel_requests", + "input_cost_per_token", + "output_cost_per_token", + "cache_read_input_token_cost", + "cache_creation_input_token_cost", + "cache_creation_input_token_cost_above_1hr", + } +) + + +class _StrictModel(BaseModel): + model_config = ConfigDict(extra="forbid", frozen=True, strict=True) + + +class _CacheControl(_StrictModel): + type: Literal["ephemeral"] + ttl: Literal["5m", "1h"] = "5m" + + +class _Text(_StrictModel): + type: Literal["text"] + text: str = Field(min_length=1, pattern=r"\S") + cache_control: _CacheControl | None = None + + +class _ToolUse(_StrictModel): + type: Literal["tool_use"] + id: str = Field(min_length=1) + name: str = Field(min_length=1) + input: Mapping[str, JsonValue] + cache_control: _CacheControl | None = None + + +class _ResultText(_StrictModel): + type: Literal["text"] + text: str + + +class _ToolResult(_StrictModel): + type: Literal["tool_result"] + tool_use_id: str = Field(min_length=1) + content: str | Annotated[tuple[_ResultText, ...], Field(strict=False)] + is_error: bool | None = None + cache_control: _CacheControl | None = None + + +_Block: TypeAlias = Annotated[_Text | _ToolUse | _ToolResult, Field(discriminator="type")] + + +class _Message(_StrictModel): + role: Literal["user", "assistant"] + content: str | Annotated[tuple[_Block, ...], Field(strict=False)] + + def blocks(self) -> tuple[_Text | _ToolUse | _ToolResult, ...]: + return (_Text(type="text", text=self.content),) if isinstance(self.content, str) else tuple(self.content) + + +class _Tool(_StrictModel): + name: str = Field(min_length=1) + description: str | None = None + input_schema: Mapping[str, JsonValue] + type: Literal["custom"] | None = None + + +class _Request(_StrictModel): + messages: tuple[_Message, ...] = Field(min_length=1, strict=False) + system: str | Annotated[tuple[_ResultText, ...], Field(strict=False)] | None = None + tools: Annotated[tuple[_Tool, ...], Field(strict=False)] | None = None + model: str | None = None + max_tokens: int | None = None + stream: bool | None = None + temperature: float | int | None = None + top_p: float | int | None = None + top_k: int | None = None + stop_sequences: Annotated[tuple[str, ...], Field(strict=False)] | None = None + metadata: Mapping[str, JsonValue] | None = None + + +@dataclass(frozen=True, slots=True) +class PromptPrefix: + prefix_body: Mapping[str, JsonValue] + fingerprint: str + fingerprints: tuple[str, ...] + ttl_seconds: int + + +def _digest(value: object) -> str: + return hashlib.sha256( + json.dumps(value, sort_keys=True, separators=(",", ":"), ensure_ascii=False).encode() + ).hexdigest() + + +def _next_digest(previous: str, boundary: tuple[int, str, Mapping[str, JsonValue]]) -> str: + return _digest((previous, boundary)) + + +def parse_prompt(body: Mapping[str, JsonValue]) -> PromptPrefix | None: + try: + request: Final = _Request.model_validate(body) + blocks: Final = tuple(message.blocks() for message in request.messages) + except ValidationError: + return None + markers: Final = tuple( + (message_index, block_index, block.cache_control) + for message_index, message_blocks in enumerate(blocks) + for block_index, block in enumerate(message_blocks) + if block.cache_control is not None + ) + if len(markers) != 1: + return None + message_end, block_end, marker = markers[0] + normalized: Final = _JSON_OBJECT.validate_python(request.model_dump(mode="json", exclude_none=True)) + context: Final = MappingProxyType({key: normalized[key] for key in ("system", "tools") if key in normalized}) + boundaries: Final = tuple( + ( + message_index, + request.messages[message_index].role, + _JSON_OBJECT.validate_python( + block.model_dump(mode="json", exclude=MappingProxyType({"cache_control": True}), exclude_none=True) + ), + ) + for message_index, message_blocks in enumerate(blocks[: message_end + 1]) + for block_index, block in enumerate(message_blocks) + if message_index < message_end or block_index <= block_end + ) + hashes: Final = tuple( + accumulate(boundaries, _next_digest, initial=_digest((_JSON_OBJECT.validate_python(context), marker.ttl))) + )[1:] + prefix_messages: Final = tuple( + _Message( + role=request.messages[message_index].role, + content=tuple( + block + for block_index, block in enumerate(message_blocks) + if message_index < message_end or block_index <= block_end + ), + ) + for message_index, message_blocks in enumerate(blocks[: message_end + 1]) + ) + return PromptPrefix( + prefix_body=MappingProxyType( + _JSON_OBJECT.validate_python( + _Request(messages=prefix_messages, system=request.system, tools=request.tools).model_dump( + mode="json", exclude_none=True + ) + ) + ), + fingerprint=hashes[-1], + fingerprints=tuple(reversed(hashes[-20:])), + ttl_seconds=3600 if marker.ttl == "1h" else 300, + ) + + +def cache_scope( + caller_key_hash: str, + deployment_id: str, + provider_key: str, + model: str, + anthropic_version: str = DEFAULT_ANTHROPIC_API_VERSION, +) -> str: + return _digest((caller_key_hash, deployment_id, provider_key, model, anthropic_version)) + + +class _TTLUsage(BaseModel): + model_config = ConfigDict(strict=True) + ephemeral_5m_input_tokens: int = Field(default=0, ge=0) + ephemeral_1h_input_tokens: int = Field(default=0, ge=0) + + +class _CacheUsage(BaseModel): + model_config = ConfigDict(strict=True) + cached_tokens: int = Field(default=0, ge=0) + cache_creation_tokens: int = Field(default=0, ge=0) + cache_creation_token_details: _TTLUsage | None = None + + +class _Usage(BaseModel): + model_config = ConfigDict(strict=True) + prompt_tokens: int = Field(ge=0) + prompt_tokens_details: _CacheUsage + + +class _Choice(BaseModel): + finish_reason: str = Field(min_length=1) + + +class _Response(BaseModel): + model_config = ConfigDict(strict=True) + model: str + usage: _Usage + choices: tuple[_Choice, ...] = Field(min_length=1, strict=False) + + +class _CountBody(BaseModel): + messages: Sequence[Mapping[str, JsonValue]] + tools: Sequence[Mapping[str, JsonValue]] | None = None + system: str | Sequence[Mapping[str, JsonValue]] | None = None + + +class _CountResult(BaseModel): + input_tokens: Annotated[StrictInt, Field(ge=0)] + + +class TokenCounter(Protocol): + async def __call__(self, model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: ... + + +def _count_objects( + values: Sequence[Mapping[str, JsonValue]], +) -> list[dict[str, JsonValue]]: # mutable-ok: the existing provider count API requires JSON lists/dicts + return [dict(value) for value in values] # mutable-ok: serialize read-only inputs at the provider API boundary + + +async def count_prompt_tokens(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + native: Final = _CountBody.model_validate(body) + try: + result: Final = _CountResult.model_validate( + await _counter.handle_count_tokens_request( + model=model, + messages=_count_objects(native.messages), + tools=_count_objects(native.tools) if native.tools is not None else None, + system=native.system, + api_key=api_key, + timeout=15.0, + ) + ) + except Exception: # noqa: BLE001 # provider/count validation failures are unavailable estimates, not zero tokens + return None + return result.input_tokens + + +@dataclass(frozen=True, slots=True) +class NativePredictionTarget: + model: str + api_key: str + + +@dataclass(frozen=True, slots=True) +class UnsupportedPredictionTarget: + reason: Literal[ + "unsupported_deployment_configuration", + "unsupported_provider_endpoint", + "unsupported_provider", + "unsupported_provider_credentials", + ] + + +def resolve_prediction_target(params: LiteLLM_Params) -> NativePredictionTarget | UnsupportedPredictionTarget: + configured_options: Final = frozenset(params.model_dump(exclude_defaults=True, exclude_none=True)) + if configured_options - _DEPLOYMENT_OPTIONS: + return UnsupportedPredictionTarget("unsupported_deployment_configuration") + api_base: Final = AnthropicModelInfo.get_api_base(params.api_base) + if api_base not in ("https://api.anthropic.com", "https://api.anthropic.com/v1/messages"): + return UnsupportedPredictionTarget("unsupported_provider_endpoint") + try: + model, provider, _, _ = litellm.get_llm_provider( + model=params.model, custom_llm_provider=params.custom_llm_provider + ) + except Exception: # noqa: BLE001 # the shared provider resolver raises for unknown deployments + return UnsupportedPredictionTarget("unsupported_provider") + if provider != "anthropic": + return UnsupportedPredictionTarget("unsupported_provider") + api_key: Final = AnthropicModelInfo.get_api_key(params.api_key) + if api_key is None or not _supported_provider_key(api_key): + return UnsupportedPredictionTarget("unsupported_provider_credentials") + return NativePredictionTarget(model=model, api_key=api_key) + + +def _supported_provider_key(api_key: str) -> bool: + return bool(api_key) and not is_anthropic_oauth_key(api_key) + + +def supported_prediction_headers(headers: Mapping[str, str]) -> bool: + return all( + name.lower() != "anthropic-beta" + and (name.lower() != "anthropic-version" or value == DEFAULT_ANTHROPIC_API_VERSION) + for name, value in headers.items() + ) + + +@dataclass(frozen=True, slots=True) +class ObservedCachePrefix: + prefix: PromptPrefix + scope: str + cached_tokens: int + cache_creation_tokens: int + + +def parse_observed_cache( + wire: httpx.Request, response_obj: ModelResponse, caller_key_hash: str, deployment_id: str +) -> ObservedCachePrefix | None: + try: + response: Final = _Response.model_validate(response_obj, from_attributes=True) + body: Final = _JSON_OBJECT.validate_json(wire.content) + headers: Final = _HEADERS.validate_python(wire.headers) + except (ValidationError, RuntimeError, httpx.RequestNotRead): + return None + if ( + wire.url.scheme != "https" + or wire.url.host != "api.anthropic.com" + or wire.url.path != "/v1/messages" + or wire.url.query + or wire.url.port not in (None, 443) + ): + return None + if ( + frozenset(headers) - _NATIVE_HEADERS + or not supported_prediction_headers(headers) + or headers.get("anthropic-version") != DEFAULT_ANTHROPIC_API_VERSION + ): + return None + provider_key: Final = headers.get("x-api-key", "") + model: Final = body.get("model") + if not _supported_provider_key(provider_key) or not isinstance(model, str) or model != response.model: + return None + prefix: Final = parse_prompt(body) + if prefix is None: + return None + usage: Final = response.usage.prompt_tokens_details + cache_tokens: Final = usage.cached_tokens + usage.cache_creation_tokens + if cache_tokens <= 0 or cache_tokens > response.usage.prompt_tokens: + return None + split: Final = usage.cache_creation_token_details + if usage.cache_creation_tokens and split is None: + return None + if split is not None and ( + split.ephemeral_5m_input_tokens + split.ephemeral_1h_input_tokens != usage.cache_creation_tokens + or (prefix.ttl_seconds == 300 and split.ephemeral_1h_input_tokens > 0) + or (prefix.ttl_seconds == 3600 and split.ephemeral_5m_input_tokens > 0) + ): + return None + return ObservedCachePrefix( + prefix=prefix, + scope=cache_scope(caller_key_hash, deployment_id, provider_key, model), + cached_tokens=cache_tokens, + cache_creation_tokens=usage.cache_creation_tokens, + ) diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index ae6c042ab3a..de82d8ec3c8 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -887,6 +887,7 @@ class LiteLLMRoutes(enum.Enum): "/auto_router/validate_complexity_router_config", # Per-session auto-router read - the endpoint scopes the row to the caller's own key hash "/auto_router/session", + "/cost/predict-cache", # Agent registry - reads are role-scoped and writes are proxy-admin-gated # inside agent_endpoints/endpoints.py *agent_management_routes, diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index 9c175242a9a..3495bf2ae98 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -878,6 +878,7 @@ async def common_checks( request_query_params=_safe_get_request_query_params(request=request), llm_router=llm_router, request=request, + team_id=valid_token.team_id if valid_token is not None else None, ) skip_all_budget_checks: Final = skip_budget_checks or ( @@ -4356,7 +4357,7 @@ async def stamp_matched_model_access_groups( async def can_key_call_model( model: str | list[str], - llm_model_list: list | None, + llm_model_list: Sequence[object] | None, valid_token: UserAPIKeyAuth, llm_router: litellm.Router | None, ) -> Literal[True]: @@ -4403,7 +4404,7 @@ async def can_key_call_model( async def can_key_call_resolved_model( model: str, - llm_model_list: list | None, + llm_model_list: Sequence[object] | None, valid_token: UserAPIKeyAuth, llm_router: litellm.Router | None, ) -> None: diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py index be65c3b39ec..dc304a156cf 100644 --- a/litellm/proxy/auth/auth_utils.py +++ b/litellm/proxy/auth/auth_utils.py @@ -33,7 +33,7 @@ from litellm.proxy.common_utils.http_parsing_utils import extract_nested_form_me from litellm.types.passthrough_endpoints.pass_through_endpoints import ( LITELLM_PASS_THROUGH_ENDPOINT_MARKER, ) -from litellm.types.router import CONFIGURABLE_CLIENTSIDE_AUTH_PARAMS +from litellm.types.router import CONFIGURABLE_CLIENTSIDE_AUTH_PARAMS, Deployment from litellm.types.utils import CustomPricingLiteLLMParams @@ -1736,7 +1736,7 @@ def _append_model_candidates(candidates: list[str], value: Any) -> None: candidates.extend(model for model in model_names if model) -def _dedupe_model_candidates(candidates: list[str]) -> list[str]: +def _dedupe_model_candidates(candidates: Collection[str]) -> list[str]: deduped: Final[list[str]] = [] for model in candidates: if model not in deduped: @@ -1845,13 +1845,42 @@ def _resolve_model_id_with_router(model_id: str | None, llm_router: Router | Non return model_id +def get_cache_prediction_deployments( + *, current_deployment_id: str, candidate_deployment_id: str, llm_router: Router, team_id: str | None +) -> tuple[Deployment, Deployment] | None: + current: Final = llm_router.get_deployment(current_deployment_id) + candidate: Final = llm_router.get_deployment(candidate_deployment_id) + if current is None or candidate is None: + return None + if any(deployment.model_info.team_id not in (None, team_id) for deployment in (current, candidate)): + return None + return current, candidate + + +def _cache_prediction_model_candidates( + request_data: Mapping[str, object], llm_router: Router | None, team_id: str | None +) -> tuple[str, ...]: + current_id: Final = request_data.get("current_deployment_id") + candidate_id: Final = request_data.get("candidate_deployment_id") + if llm_router is None or not isinstance(current_id, str) or not isinstance(candidate_id, str): + return () + deployments: Final = get_cache_prediction_deployments( + current_deployment_id=current_id, candidate_deployment_id=candidate_id, llm_router=llm_router, team_id=team_id + ) + return tuple(deployment.model_name for deployment in deployments) if deployments is not None else () + + def _extract_model_candidates_from_request( request_data: dict, route: str, request_headers: Mapping[str, object] | None = None, request_query_params: Mapping[str, object] | None = None, llm_router: Router | None = None, + team_id: str | None = None, ) -> list[str]: + if route == "/cost/predict-cache": + prediction_models: Final = _cache_prediction_model_candidates(request_data, llm_router, team_id) # pyright: ignore[reportUnknownArgumentType] # the typed reader validates each deployment ID from this legacy payload + return _dedupe_model_candidates(prediction_models) candidates: Final[list[str]] = [] uses_model_routing_sources: Final = _route_uses_model_routing_sources(route=route) uses_header_or_query_model_sources: Final = _route_matches_any_marker( @@ -1945,6 +1974,7 @@ def get_model_from_request( request_query_params: Mapping[str, object] | None = None, llm_router: Router | None = None, request: Request | None = None, + team_id: str | None = None, ) -> str | list[str] | None: """Resolve the model(s) a request targets, for model-access and budget checks. @@ -1967,6 +1997,7 @@ def get_model_from_request( request_headers=request_headers, request_query_params=request_query_params, llm_router=llm_router, + team_id=team_id, ) model = _format_model_candidates(candidates) diff --git a/litellm/proxy/auth/user_api_key_auth.py b/litellm/proxy/auth/user_api_key_auth.py index 9828311112e..930e3cca703 100644 --- a/litellm/proxy/auth/user_api_key_auth.py +++ b/litellm/proxy/auth/user_api_key_auth.py @@ -182,6 +182,7 @@ def _get_model_from_request_context( route: str, request: Request | None, llm_router: Any | None = None, + team_id: str | None = None, ) -> str | list[str] | None: return get_model_from_request( request_data=request_data, @@ -190,6 +191,7 @@ def _get_model_from_request_context( request_query_params=_safe_get_request_query_params(request=request), llm_router=llm_router, request=request, + team_id=team_id, ) @@ -208,7 +210,7 @@ async def _normalize_claude_model( return if request is not None and request.scope.get(_CLAUDE_MODEL_NORMALIZED) is True: return - requested: Final = _get_model_from_request_context(request_data, route, request, llm_router) + requested: Final = _get_model_from_request_context(request_data, route, request, llm_router, valid_token.team_id) if not isinstance(requested, str) or requested != request_data.get("model"): return if not requested.startswith("claude-router-") and not requested.lower().endswith("[1m]"): @@ -1592,6 +1594,7 @@ async def _user_api_key_auth_builder( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) skip_budget_checks = False if model is not None and llm_router is not None: @@ -1632,6 +1635,7 @@ async def _user_api_key_auth_builder( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) ), ) @@ -2022,6 +2026,7 @@ async def _user_api_key_auth_builder( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) skip_budget_checks = False if model is not None and llm_router is not None: @@ -2140,6 +2145,7 @@ async def _user_api_key_auth_builder( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) current_models = _get_model_names_for_budget_checks(model=current_model) @@ -2170,6 +2176,7 @@ async def _user_api_key_auth_builder( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) current_models = _get_model_names_for_budget_checks(model=current_model) @@ -2665,6 +2672,7 @@ async def _run_centralized_common_checks( route=route, request=request, llm_router=llm_router, + team_id=user_api_key_auth_obj.team_id, ) # Pin the metadata variable name (litellm_metadata vs metadata) before @@ -2781,12 +2789,14 @@ def _should_skip_budget_checks( route: str, request: Request | None, llm_router: Any | None, + team_id: str | None = None, ) -> bool: model: Final = _get_model_from_request_context( request_data=request_data, route=route, request=request, llm_router=llm_router, + team_id=team_id, ) if model is not None and llm_router is not None: return _is_model_cost_zero(model=model, llm_router=llm_router) @@ -3232,6 +3242,7 @@ async def _enforce_key_and_fallback_model_access( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) if model is not None: @@ -3339,6 +3350,7 @@ async def _run_post_custom_auth_checks( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) current_models = _get_model_names_for_budget_checks(model=current_model) @@ -3380,6 +3392,7 @@ async def _run_post_custom_auth_checks( route=route, request=request, llm_router=llm_router, + team_id=valid_token.team_id, ) current_models = _get_model_names_for_budget_checks(model=current_model) diff --git a/litellm/proxy/common_utils/prompt_cache_pricing.py b/litellm/proxy/common_utils/prompt_cache_pricing.py new file mode 100644 index 00000000000..ff070853b46 --- /dev/null +++ b/litellm/proxy/common_utils/prompt_cache_pricing.py @@ -0,0 +1,91 @@ +from collections.abc import Mapping +from math import isfinite +from typing import Final + +from pydantic import TypeAdapter + +import litellm +from litellm.cost_calculator import ( + _select_model_name_for_cost_calc, # pyright: ignore[reportPrivateUsage] # shares completion_cost's deployment tariff selection + completion_cost, # pyright: ignore[reportUnknownVariableType] # legacy optional parameters are untyped +) +from litellm.litellm_core_utils.litellm_logging import Logging +from litellm.types.management_endpoints.prompt_cache_prediction import CacheTokenBuckets +from litellm.types.utils import CacheCreationTokenDetails, ModelResponse, PromptTokensDetailsWrapper, Usage + +_PRICE_ENTRY: Final = TypeAdapter(Mapping[str, object]) + + +def _valid_price(value: object) -> bool: + return isinstance(value, (int, float)) and not isinstance(value, bool) and isfinite(value) and value >= 0 + + +def _has_required_prices(prices: Mapping[str, object], tokens: CacheTokenBuckets) -> bool: + required: Final = ( + ("input_cost_per_token", True), + ("cache_read_input_token_cost", tokens.cache_read_input_tokens > 0), + ("cache_creation_input_token_cost", tokens.cache_creation_5m_input_tokens > 0), + ("cache_creation_input_token_cost_above_1hr", tokens.cache_creation_1h_input_tokens > 0), + ) + if any(needed and not _valid_price(prices.get(key)) for key, needed in required): + return False + return all( + _valid_price(value) + for key, value in prices.items() + if value is not None and any(needed and key.startswith(f"{base}_above_") for base, needed in required) + ) + + +def price_cache_tokens(model: str, deployment_id: str, tokens: CacheTokenBuckets) -> float | None: + try: + selected_model: Final = _select_model_name_for_cost_calc( + model=model, + completion_response=None, + custom_pricing=True, + custom_llm_provider="anthropic", + router_model_id=deployment_id, + ) + if selected_model is None: + return None + model_info: Final = litellm.get_model_info(model=selected_model, custom_llm_provider="anthropic") + registry: Final = _PRICE_ENTRY.validate_python(litellm.model_cost) # pyright: ignore[reportUnknownMemberType] # legacy registry is validated at this boundary + price_entry: Final = registry.get(model_info["key"]) + if price_entry is None: + return None + prices: Final = _PRICE_ENTRY.validate_python(price_entry) + if not _has_required_prices(prices, tokens): + return None + usage: Final = Usage( + prompt_tokens=tokens.total_tokens, + completion_tokens=0, + total_tokens=tokens.total_tokens, + prompt_tokens_details=PromptTokensDetailsWrapper( + cached_tokens=tokens.cache_read_input_tokens, + cache_creation_tokens=tokens.cache_creation_5m_input_tokens + tokens.cache_creation_1h_input_tokens, + cache_creation_token_details=CacheCreationTokenDetails( + ephemeral_5m_input_tokens=tokens.cache_creation_5m_input_tokens, + ephemeral_1h_input_tokens=tokens.cache_creation_1h_input_tokens, + ), + ), + ) + logging_obj: Final = Logging( + model=model, + messages=[], # mutable-ok: Logging requires a list + stream=False, + call_type="completion", + start_time=None, + litellm_call_id="prompt-cache-prediction", + function_id="prompt-cache-prediction", + ) + completion_cost( + completion_response=ModelResponse(model=model, usage=usage), + model=model, + custom_llm_provider="anthropic", + custom_pricing=True, + router_model_id=deployment_id, + litellm_logging_obj=logging_obj, + ) + cost: Final = logging_obj.cost_breakdown.get("input_cost") if logging_obj.cost_breakdown is not None else None + return cost if cost is not None and _valid_price(cost) else None + except Exception: # noqa: BLE001 # the shared pricing owners raise plain Exception for unpriceable models + return None diff --git a/litellm/proxy/hooks/__init__.py b/litellm/proxy/hooks/__init__.py index 8714dd5f3d2..f3542098f95 100644 --- a/litellm/proxy/hooks/__init__.py +++ b/litellm/proxy/hooks/__init__.py @@ -9,6 +9,7 @@ from .max_budget_per_session_limiter import _PROXY_MaxBudgetPerSessionHandler from .max_iterations_limiter import _PROXY_MaxIterationsHandler from .parallel_request_limiter import _PROXY_MaxParallelRequestsHandler from .parallel_request_limiter_v3 import _PROXY_MaxParallelRequestsHandler_v3 +from .prompt_cache_prediction import PromptCacheObserver from .responses_id_security import ResponsesIDSecurity from .sensitive_data_routing import _PROXY_SensitiveDataRoutingHandler @@ -25,6 +26,7 @@ PROXY_HOOKS: Final = { "max_iterations_limiter": _PROXY_MaxIterationsHandler, "max_budget_per_session_limiter": _PROXY_MaxBudgetPerSessionHandler, "sensitive_data_routing": _PROXY_SensitiveDataRoutingHandler, + "prompt_cache_prediction": PromptCacheObserver, } ## FEATURE FLAG HOOKS ## diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py index c398abff099..a34dc99e472 100644 --- a/litellm/proxy/hooks/parallel_request_limiter_v3.py +++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py @@ -9,10 +9,12 @@ import binascii import logging import os import uuid -from collections.abc import Awaitable, Callable, Mapping, Sequence, Set +from collections.abc import AsyncGenerator, Awaitable, Callable, Mapping, Sequence, Set +from contextlib import asynccontextmanager from contextvars import ContextVar from dataclasses import dataclass, field from datetime import datetime +from types import MappingProxyType from typing import ( TYPE_CHECKING, Any, @@ -23,6 +25,7 @@ from typing import ( TypedDict, ) +from pydantic import TypeAdapter from typing_extensions import NotRequired, ReadOnly from litellm import DualCache @@ -84,6 +87,9 @@ else: InternalUsageCache = Any +_REQUEST_RATE_LIMIT_DATA: Final = TypeAdapter(Mapping[str, object]) + + BATCH_RATE_LIMITER_SCRIPT: Final = """ local results = {} local now = tonumber(ARGV[1]) @@ -2673,12 +2679,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): Returns list of descriptors for API key, user, team, team member, end user, model-specific, agent, and agent-session limits. """ - from litellm.proxy.auth.auth_utils import ( - get_team_model_rpm_limit, - get_team_model_tpm_limit, - ) - - descriptors: Final = [] + descriptors: Final[list[RateLimitDescriptor]] = [] # mutable-ok: existing descriptor helpers append in place # API Key rate limits if user_api_key_dict.api_key and ( @@ -2803,34 +2804,11 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): descriptors=descriptors, ) - if ( - get_team_model_rpm_limit(user_api_key_dict) is not None - or get_team_model_tpm_limit(user_api_key_dict) is not None - ): - _tpm_limit_for_team_model: Final = get_team_model_tpm_limit(user_api_key_dict) or {} - _rpm_limit_for_team_model: Final = get_team_model_rpm_limit(user_api_key_dict) or {} - should_check_rate_limit = False - if requested_model in _tpm_limit_for_team_model or requested_model in _rpm_limit_for_team_model: - should_check_rate_limit = True - - if should_check_rate_limit: - model_specific_tpm_limit = None - model_specific_rpm_limit = None - if requested_model in _tpm_limit_for_team_model: - model_specific_tpm_limit = _tpm_limit_for_team_model[requested_model] - if requested_model in _rpm_limit_for_team_model: - model_specific_rpm_limit = _rpm_limit_for_team_model[requested_model] - descriptors.append( - RateLimitDescriptor( - key="model_per_team", - value=f"{user_api_key_dict.team_id}:{requested_model}", - rate_limit={ - "requests_per_unit": model_specific_rpm_limit, - "tokens_per_unit": model_specific_tpm_limit, - "window_size": self.window_size, - }, - ) - ) + self._add_team_model_rate_limit_descriptor_from_metadata( + user_api_key_dict=user_api_key_dict, + requested_model=requested_model if isinstance(requested_model, str) else None, + descriptors=descriptors, + ) # Agent-level and session-level rate limits resolved_agent_id: Final = self._get_resolved_agent_id(user_api_key_dict, data) @@ -3416,6 +3394,108 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): requested_model, ) + async def _build_request_rate_limit_descriptors( + self, + user_api_key_dict: UserAPIKeyAuth, + data: Mapping[str, object], + call_type: str | None, + ) -> list[RateLimitDescriptor]: # mutable-ok: the shared generation reservation helpers require a list + metadata: Final = _REQUEST_RATE_LIMIT_DATA.validate_python( + user_api_key_dict.metadata or MappingProxyType({}) # pyright: ignore[reportUnknownMemberType] # validates the legacy auth metadata boundary + ) + rpm_value: Final = metadata.get("rpm_limit_type") + tpm_value: Final = metadata.get("tpm_limit_type") + rpm_limit_type: Final = rpm_value if isinstance(rpm_value, str) else None + tpm_limit_type: Final = tpm_value if isinstance(tpm_value, str) else None + model_value: Final = data.get("model") + requested_model: Final = model_value if isinstance(model_value, str) else None + model_has_failures: Final = ( + await self._check_model_has_recent_failures( + model=requested_model, + parent_otel_span=user_api_key_dict.parent_otel_span, + ) + if requested_model and self._is_dynamic_rate_limiting_enabled(rpm_limit_type, tpm_limit_type) + else False + ) + descriptors: Final = self._create_rate_limit_descriptors( # pyright: ignore[reportUnknownMemberType] # legacy helper reads a dictionary with validated keys + user_api_key_dict=user_api_key_dict, + data=dict(data), # mutable-ok: legacy descriptor helpers accept a request dictionary + rpm_limit_type=rpm_limit_type, + tpm_limit_type=tpm_limit_type, + model_has_failures=model_has_failures, + call_type=call_type, + ) + self._add_project_model_rate_limit_descriptor_from_metadata( + user_api_key_dict=user_api_key_dict, + requested_model=requested_model, + descriptors=descriptors, + ) + self.add_project_io_token_rate_limit_descriptors_from_metadata( + user_api_key_dict=user_api_key_dict, + requested_model=requested_model, + descriptors=descriptors, + ) + return [ # mutable-ok: the shared generation reservation helpers require a list + *descriptors, + *self.create_organization_rate_limit_descriptor(user_api_key_dict, requested_model), + ] + + async def _release_request_capacity_when_admitted( + self, + admission: asyncio.Task[RateLimitResponse], + acquisition: ParallelSlotAcquisition, + user_api_key_dict: UserAPIKeyAuth, + ) -> None: + response: Final = await admission + if response["overall_code"] == "OK": + await self._release_parallel_request_slots(acquisition, user_api_key_dict.parent_otel_span) + + @asynccontextmanager + async def request_capacity( + self, + user_api_key_dict: UserAPIKeyAuth, + model: str, + *, + request_data: Mapping[str, object] | None = None, + ) -> AsyncGenerator[None, None]: + """Charge one non-generation provider request to RPM and hold its concurrency slot.""" + data: Final = MappingProxyType({**(request_data or MappingProxyType({})), "model": model}) + descriptors: Final = await self._build_request_rate_limit_descriptors(user_api_key_dict, data, None) + acquisition: Final = ParallelSlotAcquisition( + slot_id=uuid.uuid4().hex, + counter_keys=[ # mutable-ok: the shared slot-release contract requires a list + self.create_rate_limit_keys(d["key"], d["value"], "max_parallel_requests") + for d in descriptors + if d["rate_limit"] is not None and d["rate_limit"].get("max_parallel_requests") is not None + ], + ) + admission: Final = asyncio.create_task( + self.should_rate_limit( + descriptors=descriptors, + parent_otel_span=user_api_key_dict.parent_otel_span, + skip_tpm_check=True, + parallel_slot_id=acquisition["slot_id"], + ) + ) + try: + response: Final = await asyncio.shield(admission) + if response["overall_code"] == "OVER_LIMIT": + self._handle_rate_limit_error(response, descriptors, model) + yield + finally: + cleanup: Final = asyncio.create_task( + self._release_request_capacity_when_admitted(admission, acquisition, user_api_key_dict) + ) + cancellation: asyncio.CancelledError | None = None # rebind-ok: retain cancellation until cleanup finishes + while not cleanup.done(): + try: + await asyncio.shield(cleanup) + except asyncio.CancelledError as exc: + cancellation = exc # rebind-ok: retain the latest cancellation without interrupting slot release + cleanup.result() + if cancellation is not None: + raise cancellation + async def async_pre_call_hook( self, user_api_key_dict: UserAPIKeyAuth, @@ -3444,59 +3524,15 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): call_type=call_type, ) - # Get rate limit types from metadata - metadata: Final = user_api_key_dict.metadata or {} - rpm_limit_type: Final = metadata.get("rpm_limit_type") - tpm_limit_type: Final = metadata.get("tpm_limit_type") - - # For dynamic mode, check if the model has recent failures - model_has_failures = False - requested_model: Final = data.get("model", None) - - if ( - self._is_dynamic_rate_limiting_enabled( - rpm_limit_type=rpm_limit_type, - tpm_limit_type=tpm_limit_type, - ) - and requested_model - ): - model_has_failures = await self._check_model_has_recent_failures( - model=requested_model, - parent_otel_span=user_api_key_dict.parent_otel_span, - ) - - # Create rate limit descriptors - descriptors: Final = self._create_rate_limit_descriptors( + request_data: Final = _REQUEST_RATE_LIMIT_DATA.validate_python(data) + model_value: Final = request_data.get("model") + requested_model: Final = model_value if isinstance(model_value, str) else None + descriptors: Final = await self._build_request_rate_limit_descriptors( user_api_key_dict=user_api_key_dict, - data=data, - rpm_limit_type=rpm_limit_type, - tpm_limit_type=tpm_limit_type, - model_has_failures=model_has_failures, + data=request_data, call_type=call_type, ) - # Add team model rate limits from team_metadata - self._add_team_model_rate_limit_descriptor_from_metadata( - user_api_key_dict=user_api_key_dict, - requested_model=requested_model, - descriptors=descriptors, - ) - - # Project Level Rate Limits - self._add_project_model_rate_limit_descriptor_from_metadata( - user_api_key_dict=user_api_key_dict, - requested_model=requested_model, - descriptors=descriptors, - ) - self.add_project_io_token_rate_limit_descriptors_from_metadata( - user_api_key_dict=user_api_key_dict, - requested_model=requested_model, - descriptors=descriptors, - ) - - # Org Level Rate Limits - descriptors.extend(self.create_organization_rate_limit_descriptor(user_api_key_dict, requested_model)) - # Only check rate limits if we have descriptors with actual limits if descriptors: # First pass: RPM and max_parallel_requests sliding-window check. diff --git a/litellm/proxy/hooks/prompt_cache_prediction.py b/litellm/proxy/hooks/prompt_cache_prediction.py new file mode 100644 index 00000000000..65c456c5666 --- /dev/null +++ b/litellm/proxy/hooks/prompt_cache_prediction.py @@ -0,0 +1,142 @@ +from __future__ import annotations + +import asyncio +import time +from collections.abc import Callable, Mapping +from datetime import datetime +from typing import TYPE_CHECKING, Final, Literal + +import httpx +from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError + +from litellm.caching.dual_cache import DualCache +from litellm.integrations.custom_logger import CustomLogger +from litellm.llms.anthropic.prompt_cache_prediction import PromptPrefix, parse_observed_cache +from litellm.types.utils import ModelResponse + +if TYPE_CHECKING: + from litellm.proxy.utils import InternalUsageCache + +_RETENTION_SECONDS: Final = 86_400 + + +class CacheObservation(BaseModel): + model_config = ConfigDict(extra="forbid", frozen=True, strict=True) + + fingerprint: str = Field(pattern=r"^[0-9a-f]{64}$") + cached_tokens: int = Field(gt=0) + observed_at: float = Field(ge=0, allow_inf_nan=False) + expires_at: float = Field(ge=0, allow_inf_nan=False) + + +_CACHE_ENTRY: Final[TypeAdapter[CacheObservation | str | None]] = TypeAdapter(CacheObservation | str | None) + + +def _cache_key(scope: str, fingerprint: str) -> str: + return f"prompt-cache-observation:{scope}:{fingerprint}" + + +async def lookup( + cache: DualCache, scope: str, prefix: PromptPrefix, now: float | None = None +) -> CacheObservation | None: + checked_at: Final = time.time() if now is None else now + exact: Final = await _read_exact(cache, scope, prefix.fingerprint) + if exact is not None and exact.expires_at > checked_at: + return exact + older: Final = await asyncio.gather( + *(_read_exact(cache, scope, fingerprint) for fingerprint in prefix.fingerprints[1:]) + ) + observations: Final = tuple(observation for observation in (exact, *older) if observation is not None) + return next( + (observation for observation in observations if observation.expires_at > checked_at), + next(iter(observations), None), + ) + + +async def _read_exact(cache: DualCache, scope: str, fingerprint: str) -> CacheObservation | None: + try: + value: Final = _CACHE_ENTRY.validate_python(await cache.async_get_cache(_cache_key(scope, fingerprint), ttl=1)) # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType] # validate the legacy cache's untyped result at the I/O boundary + if value is None: + return None + observation: Final = CacheObservation.model_validate_json(value) if isinstance(value, str) else value + except ValidationError: + return None + return observation if observation.fingerprint == fingerprint else None + + +class _Metadata(BaseModel): + model_config = ConfigDict(strict=True) + user_api_key_hash: str = Field(min_length=1) + + +class _Logged(BaseModel): + model_config = ConfigDict(strict=True) + status: Literal["success"] + model_id: str = Field(min_length=1) + metadata: _Metadata + + +class _Event(BaseModel): + model_config = ConfigDict(strict=True, arbitrary_types_allowed=True) + call_type: Literal["anthropic_messages"] + custom_llm_provider: Literal["anthropic"] + cache_hit: bool | None = None + httpx_response: httpx.Response + first_api_call_start_time: datetime + standard_logging_object: _Logged + stream: bool = False + prompt_cache_response_complete: bool = False + + +class PromptCacheObserver(CustomLogger): + def __init__(self, internal_usage_cache: InternalUsageCache, clock: Callable[[], float] = time.time) -> None: + super().__init__() # pyright: ignore[reportUnknownMemberType] # base callback constructor accepts untyped kwargs + self.cache = internal_usage_cache.dual_cache + self.clock = clock + + async def async_log_success_event( + self, kwargs: Mapping[str, object], response_obj: object, start_time: datetime, end_time: datetime + ) -> None: + if not isinstance(response_obj, ModelResponse): + return + try: + event: Final = _Event.model_validate(kwargs) + wire: Final = event.httpx_response.request + except (ValidationError, RuntimeError, httpx.RequestNotRead): + return + if ( + event.cache_hit + or event.httpx_response.status_code != 200 + or (event.stream and not event.prompt_cache_response_complete) + ): + return + observed: Final = parse_observed_cache( + wire, + response_obj, + event.standard_logging_object.metadata.user_api_key_hash, + event.standard_logging_object.model_id, + ) + if observed is None: + return + prefix: Final = observed.prefix + scope: Final = observed.scope + cache_tokens: Final = observed.cached_tokens + now: Final = self.clock() + started: Final = event.first_api_call_start_time.timestamp() + if started > now: + return + if observed.cache_creation_tokens == 0: + previous: Final = await _read_exact(self.cache, scope, prefix.fingerprint) + if previous is None or previous.fingerprint != prefix.fingerprint or previous.cached_tokens != cache_tokens: + return + observation: Final = CacheObservation( + fingerprint=prefix.fingerprint, + cached_tokens=cache_tokens, + observed_at=now, + expires_at=started + prefix.ttl_seconds, + ) + key: Final = _cache_key(scope, prefix.fingerprint) + payload: Final = observation.model_dump_json() + await self.cache.async_set_cache(key, payload, ttl=_RETENTION_SECONDS) # pyright: ignore[reportUnknownMemberType] # legacy cache accepts a serialized validated observation + if self.cache.redis_cache is not None: + await self.cache.async_set_cache(key, payload, local_only=True, ttl=1) # pyright: ignore[reportUnknownMemberType] # keep the local copy short-lived while Redis retains stale evidence diff --git a/litellm/proxy/management_endpoints/cost_tracking_settings.py b/litellm/proxy/management_endpoints/cost_tracking_settings.py index dc0da63555f..cb376f286ec 100644 --- a/litellm/proxy/management_endpoints/cost_tracking_settings.py +++ b/litellm/proxy/management_endpoints/cost_tracking_settings.py @@ -28,6 +28,7 @@ from litellm.proxy._types import ( UserAPIKeyAuth, ) from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.proxy.management_endpoints.prompt_cache_prediction import router as prompt_cache_prediction_router from litellm.types.utils import ( CostBreakdown, CostPerToken, @@ -39,6 +40,7 @@ from litellm.types.utils import ( ) router: Final = APIRouter() +router.include_router(prompt_cache_prediction_router) @dataclass(frozen=True, slots=True) diff --git a/litellm/proxy/management_endpoints/prompt_cache_prediction.py b/litellm/proxy/management_endpoints/prompt_cache_prediction.py new file mode 100644 index 00000000000..56e844214d6 --- /dev/null +++ b/litellm/proxy/management_endpoints/prompt_cache_prediction.py @@ -0,0 +1,278 @@ +import time +from collections.abc import Mapping +from types import MappingProxyType +from typing import Annotated, Final + +from fastapi import APIRouter, Depends, HTTPException, Request +from pydantic import BaseModel, JsonValue, TypeAdapter + +import litellm +from litellm._internal_context import current_billing_time, pinned_billing_time +from litellm.caching.caching import DualCache +from litellm.integrations.custom_logger import CustomLogger +from litellm.llms.anthropic.prompt_cache_prediction import ( + PromptPrefix, + TokenCounter, + UnsupportedPredictionTarget, + cache_scope, + count_prompt_tokens, + parse_prompt, + resolve_prediction_target, + supported_prediction_headers, +) +from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.auth.auth_checks import can_key_call_resolved_model +from litellm.proxy.auth.auth_utils import get_cache_prediction_deployments +from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.proxy.common_utils.http_parsing_utils import ( + _read_request_body, # pyright: ignore[reportPrivateUsage, reportUnknownVariableType] # canonical parsed-body owner; validate its legacy result at the endpoint boundary +) +from litellm.proxy.common_utils.prompt_cache_pricing import price_cache_tokens +from litellm.proxy.hooks.parallel_request_limiter_v3 import ( + _PROXY_MaxParallelRequestsHandler_v3, # pyright: ignore[reportPrivateUsage] # use the configured proxy limiter's shared capacity owner +) +from litellm.proxy.hooks.prompt_cache_prediction import lookup +from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup +from litellm.types.management_endpoints.prompt_cache_prediction import ( + CacheCostScenario, + CacheEvidence, + CachePredictionArm, + CachePredictionRequest, + CachePredictionResponse, + CacheTokenBuckets, +) +from litellm.types.router import Deployment +from litellm.utils import get_prompt_cache_min_tokens + +router: Final = APIRouter() +_REQUEST_DATA: Final = TypeAdapter(Mapping[str, object]) + + +class _CallerSettings(BaseModel): + config: Mapping[str, object] | None = None + + +def has_request_transforms() -> bool: + from litellm.proxy.hooks import PROXY_HOOKS + + builtins: Final = frozenset(PROXY_HOOKS.values()) + hooks: Final = ("async_pre_call_hook", "async_pre_request_hook", "async_pre_call_deployment_hook") + callbacks: Final = litellm.logging_callback_manager.get_custom_loggers_for_type(callback_type=CustomLogger) + return any( + type(callback) not in builtins + and any(getattr(type(callback), hook) is not getattr(CustomLogger, hook) for hook in hooks) + for callback in callbacks + ) + + +def _buckets(prefix_tokens: int, suffix_tokens: int, read_tokens: int, ttl_seconds: int) -> CacheTokenBuckets: + return CacheTokenBuckets( + uncached_input_tokens=suffix_tokens, + cache_read_input_tokens=read_tokens, + cache_creation_5m_input_tokens=prefix_tokens - read_tokens if ttl_seconds == 300 else 0, + cache_creation_1h_input_tokens=prefix_tokens - read_tokens if ttl_seconds == 3600 else 0, + ) + + +def _scenario(model: str, deployment_id: str, tokens: CacheTokenBuckets) -> CacheCostScenario | None: + cost: Final = price_cache_tokens(model=model, deployment_id=deployment_id, tokens=tokens) + return CacheCostScenario(tokens=tokens, input_cost=cost) if cost is not None else None + + +def _capacity_counter( + limiter: _PROXY_MaxParallelRequestsHandler_v3, + caller: UserAPIKeyAuth, + model_name: str, + request_data: Mapping[str, object], +) -> TokenCounter: + async def count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + async with limiter.request_capacity(caller, model_name, request_data=request_data): + return await count_prompt_tokens(model, api_key, body) + + return count + + +def _capacity_request_data( + http_request: Request, caller: UserAPIKeyAuth, request_data: Mapping[str, object] +) -> Mapping[str, object]: + # The parsed-body cache retains only original top-level keys. Replay the + # shared idempotent tag merges on limiter-only data when auth added metadata. + data: Final = dict(request_data) # mutable-ok: the existing tag merge owners accept a dictionary out-param + LiteLLMProxyRequestSetup.apply_client_tag_policy_pre_auth(http_request, data, caller) # pyright: ignore[reportUnknownMemberType] # legacy tag owner takes the validated capacity dictionary + LiteLLMProxyRequestSetup.apply_key_tags_pre_auth(data, caller) # pyright: ignore[reportUnknownMemberType] # legacy tag owner merges trusted key tags into capacity metadata + return MappingProxyType(data) + + +async def predict_arm( + deployment: Deployment, + body: Mapping[str, JsonValue], + prefix: PromptPrefix, + caller_key_hash: str, + cache: DualCache, + token_counter: TokenCounter, +) -> CachePredictionArm: + deployment_id: Final = deployment.model_info.id or "" + params: Final = deployment.litellm_params + unknown: Final = CachePredictionArm(deployment_id=deployment_id, model=params.model) + if deployment.model_info.blocked: + return unknown.model_copy(update=MappingProxyType({"reason": "unsupported_deployment_configuration"})) + target: Final = resolve_prediction_target(params) + if isinstance(target, UnsupportedPredictionTarget): + return unknown.model_copy(update=MappingProxyType({"reason": target.reason})) + model: Final = target.model + api_key: Final = target.api_key + total_count: Final = await token_counter(model, api_key, body) + prefix_count: Final = await token_counter(model, api_key, prefix.prefix_body) + if total_count is None or prefix_count is None or total_count < prefix_count: + return unknown.model_copy(update=MappingProxyType({"reason": "token_count_unavailable"})) + scope: Final = cache_scope(caller_key_hash, deployment_id, api_key, model) + observation: Final = await lookup(cache, scope, prefix) + exact: Final = observation is not None and observation.fingerprint == prefix.fingerprint + cacheable: Final = observation.cached_tokens if exact and observation is not None else prefix_count + if cacheable > total_count or (observation is not None and observation.cached_tokens > cacheable): + return unknown.model_copy(update=MappingProxyType({"reason": "inconsistent_prefix_token_count"})) + suffix: Final = total_count - cacheable + evidence: Final = ( + CacheEvidence(observed_at=observation.observed_at, expires_at=observation.expires_at) + if observation is not None + else None + ) + if cacheable < get_prompt_cache_min_tokens(params.model): + disabled: Final = _scenario(model, deployment_id, CacheTokenBuckets(uncached_input_tokens=total_count)) + if disabled is None: + return unknown.model_copy(update=MappingProxyType({"reason": "pricing_unavailable"})) + return CachePredictionArm( + deployment_id=deployment_id, + model=model, + cache_state="disabled", + reason="below_cache_minimum", + estimate=disabled, + cold=disabled, + warm=disabled, + token_count_source="anthropic_count_tokens", + ) + fresh: Final = observation is not None and observation.expires_at > time.time() + read: Final = observation.cached_tokens if fresh and observation is not None else 0 + with pinned_billing_time(current_billing_time()): + cold: Final = _scenario(model, deployment_id, _buckets(cacheable, suffix, 0, prefix.ttl_seconds)) + warm: Final = _scenario(model, deployment_id, _buckets(cacheable, suffix, cacheable, prefix.ttl_seconds)) + estimate: Final = _scenario(model, deployment_id, _buckets(cacheable, suffix, read, prefix.ttl_seconds)) + if cold is None or warm is None or estimate is None: + return unknown.model_copy(update=MappingProxyType({"reason": "pricing_unavailable"})) + return CachePredictionArm( + deployment_id=deployment_id, + model=model, + cache_state="warm" if fresh and exact else "partial" if fresh else "stale" if observation else "unknown", + reason=None if fresh else "observation_expired" if observation else "no_compatible_observation", + estimate=estimate, + cold=cold, + warm=warm, + evidence=evidence, + token_count_source="anthropic_count_tokens", + ) + + +@router.post( + "/cost/predict-cache", + tags=["Cost Tracking"], # mutable-ok: FastAPI requires a list for OpenAPI tags + response_model=CachePredictionResponse, +) +async def predict_cache_cost( + request: CachePredictionRequest, + http_request: Request, + user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)], +) -> CachePredictionResponse: + """Compare the next native Anthropic request on two configured deployment IDs. + + Estimates use provider token counting and recent successful cache telemetry for this key. + Unknown cache state uses the cold scenario when prices/counts are available. Cache observations + do not guarantee retention. v0 supports one message-content breakpoint, text and client tools; + system/tool-only breakpoints, thinking, images, nondefault Anthropic versions, beta headers and + request transforms are unknown. + Each provider count consumes one RPM unit and holds concurrency capacity; a comparison uses + up to four counts. The legacy rate limiter returns unknown without contacting the provider. + This endpoint does not generate tokens, prewarm caches, choose a model or alter routing. + """ + from litellm.proxy.proxy_server import llm_router, proxy_logging_obj + + if llm_router is None: + raise HTTPException(status_code=503, detail="Model router is unavailable") + deployments: Final = get_cache_prediction_deployments( + current_deployment_id=request.current_deployment_id, + candidate_deployment_id=request.candidate_deployment_id, + llm_router=llm_router, + team_id=user_api_key_dict.team_id, + ) + if deployments is None: + raise HTTPException(status_code=404, detail="Deployment not found") + current, candidate = deployments + for deployment in (current, candidate): + await can_key_call_resolved_model( + model=deployment.model_name, + llm_model_list=llm_router.get_model_list(), + valid_token=user_api_key_dict, + llm_router=llm_router, + ) + prefix: Final = parse_prompt(request.request) + caller: Final = user_api_key_dict.api_key + caller_settings: Final = _CallerSettings.model_validate(user_api_key_dict, from_attributes=True) + unsupported_transform: Final = bool(caller_settings.config) or has_request_transforms() + unsupported_headers: Final = not supported_prediction_headers(http_request.headers) + limiter: Final = proxy_logging_obj.get_proxy_hook("parallel_request_limiter") + if ( + prefix is None + or not caller + or unsupported_transform + or unsupported_headers + or not isinstance(limiter, _PROXY_MaxParallelRequestsHandler_v3) + ): + reason: Final = ( + "unsupported_provider_headers" + if unsupported_headers + else "unsupported_request_transform" + if unsupported_transform + else "unsupported_prompt_shape" + if prefix is None + else "caller_identity_unavailable" + if not caller + else "limiter_unavailable" + ) + return CachePredictionResponse( + stay=CachePredictionArm(deployment_id=request.current_deployment_id, reason=reason), + switch=CachePredictionArm(deployment_id=request.candidate_deployment_id, reason=reason), + switch_delta=None, + cache_rebuild_penalty=None, + ) + request_data: Final = _capacity_request_data( + http_request, user_api_key_dict, _REQUEST_DATA.validate_python(await _read_request_body(http_request)) + ) + stay: Final = await predict_arm( + current, + request.request, + prefix, + caller, + proxy_logging_obj.internal_usage_cache.dual_cache, + _capacity_counter(limiter, user_api_key_dict, current.model_name, request_data), + ) + switch: Final = ( + stay + if current.model_info.id == candidate.model_info.id + else await predict_arm( + candidate, + request.request, + prefix, + caller, + proxy_logging_obj.internal_usage_cache.dual_cache, + _capacity_counter(limiter, user_api_key_dict, candidate.model_name, request_data), + ) + ) + return CachePredictionResponse( + stay=stay, + switch=switch, + switch_delta=(switch.estimate.input_cost - stay.estimate.input_cost) + if switch.estimate is not None and stay.estimate is not None + else None, + cache_rebuild_penalty=(switch.estimate.input_cost - switch.warm.input_cost) + if switch.estimate is not None and switch.warm is not None + else None, + ) diff --git a/litellm/proxy/pass_through_endpoints/streaming_handler.py b/litellm/proxy/pass_through_endpoints/streaming_handler.py index 4be0235adbb..b310fc661c4 100644 --- a/litellm/proxy/pass_through_endpoints/streaming_handler.py +++ b/litellm/proxy/pass_through_endpoints/streaming_handler.py @@ -270,6 +270,24 @@ class PassThroughStreamingHandler: - Vertex AI - OpenAI """ + from litellm.llms.anthropic.experimental_pass_through.messages.streaming_iterator import ( + _is_message_stop_chunk, # pyright: ignore[reportPrivateUsage] # both native stream paths share terminal-event detection + _is_provider_error_chunk, # pyright: ignore[reportPrivateUsage] # provider errors must not become cache evidence + ) + + # Transport reads can split event names and JSON payloads. Recognize terminal + # events only after the shared SSE framer has reassembled the collected bytes. + complete_frames, incomplete_tail = split_complete_sse_frames( + b"".join(raw_bytes) if endpoint_type == EndpointType.ANTHROPIC else b"" + ) + litellm_logging_obj.model_call_details[ # rebind-ok: stamp evidence on the per-request state read by callbacks + "prompt_cache_response_complete" + ] = ( + endpoint_type == EndpointType.ANTHROPIC + and not incomplete_tail.strip() + and _is_message_stop_chunk(complete_frames) + and not _is_provider_error_chunk(complete_frames) + ) try: ( standard_logging_response_object, diff --git a/litellm/types/management_endpoints/prompt_cache_prediction.py b/litellm/types/management_endpoints/prompt_cache_prediction.py new file mode 100644 index 00000000000..3789607b021 --- /dev/null +++ b/litellm/types/management_endpoints/prompt_cache_prediction.py @@ -0,0 +1,67 @@ +from collections.abc import Mapping +from typing import Annotated, Literal, TypeAlias + +from pydantic import BaseModel, ConfigDict, Field, JsonValue, StrictInt + +TokenCount: TypeAlias = Annotated[StrictInt, Field(ge=0)] + + +class CacheTokenBuckets(BaseModel): + model_config = ConfigDict(frozen=True, extra="forbid") + + uncached_input_tokens: TokenCount = 0 + cache_read_input_tokens: TokenCount = 0 + cache_creation_5m_input_tokens: TokenCount = 0 + cache_creation_1h_input_tokens: TokenCount = 0 + + @property + def total_tokens(self) -> int: + return ( + self.uncached_input_tokens + + self.cache_read_input_tokens + + self.cache_creation_5m_input_tokens + + self.cache_creation_1h_input_tokens + ) + + +class CacheEvidence(BaseModel): + model_config = ConfigDict(frozen=True) + + observed_at: float + expires_at: float + source: Literal["provider_usage"] = "provider_usage" + confidence: Literal["observed"] = "observed" + + +class CacheCostScenario(BaseModel): + tokens: CacheTokenBuckets + input_cost: float + + +class CachePredictionArm(BaseModel): + deployment_id: str + model: str | None = None + cache_state: Literal["warm", "partial", "stale", "unknown", "disabled"] = "unknown" + reason: str | None = None + estimate: CacheCostScenario | None = None + cold: CacheCostScenario | None = None + warm: CacheCostScenario | None = None + evidence: CacheEvidence | None = None + token_count_source: Literal["anthropic_count_tokens"] | None = None + + +class CachePredictionRequest(BaseModel): + model_config = ConfigDict(extra="forbid") + + current_deployment_id: str = Field(min_length=1, max_length=256) + candidate_deployment_id: str = Field(min_length=1, max_length=256) + request: Mapping[str, JsonValue] + + +class CachePredictionResponse(BaseModel): + stay: CachePredictionArm + switch: CachePredictionArm + switch_delta: float | None + cache_rebuild_penalty: float | None + pricing_basis: Literal["input_before_discounts_and_margins"] = "input_before_discounts_and_margins" + cache_guarantee: Literal[False] = False diff --git a/tests/test_litellm/llms/anthropic/test_anthropic_prompt_cache_prediction.py b/tests/test_litellm/llms/anthropic/test_anthropic_prompt_cache_prediction.py new file mode 100644 index 00000000000..2b36866a1a0 --- /dev/null +++ b/tests/test_litellm/llms/anthropic/test_anthropic_prompt_cache_prediction.py @@ -0,0 +1,209 @@ +import json +from collections.abc import Mapping +from datetime import datetime +from types import SimpleNamespace +from typing import Final + +import httpx +import pytest +import respx +from pydantic import JsonValue + +import litellm +from litellm.caching.dual_cache import DualCache +from litellm.caching.llm_caching_handler import LLMClientCache +from litellm.llms.anthropic.count_tokens import handler as count_handler +from litellm.llms.anthropic.experimental_pass_through.messages.transformation import DEFAULT_ANTHROPIC_API_VERSION +from litellm.llms.anthropic.prompt_cache_prediction import ( + NativePredictionTarget, + cache_scope, + count_prompt_tokens, + parse_observed_cache, + parse_prompt, + resolve_prediction_target, + supported_prediction_headers, +) +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler +from litellm.models.credentials import CredentialItem +from litellm.proxy import proxy_server +from litellm.proxy.hooks.prompt_cache_prediction import PromptCacheObserver, lookup +from litellm.proxy.management_endpoints.prompt_cache_prediction import predict_arm +from litellm.proxy.utils import InternalUsageCache +from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo +from litellm.types.utils import CacheCreationTokenDetails, ModelResponse, PromptTokensDetailsWrapper, Usage + +_MODEL: Final = "claude-sonnet-5" +_KEY: Final = "test-provider-key" +_CALLER: Final = "test-caller-hash" +_DEPLOYMENT: Final = "test-native-deployment" + + +def _body() -> dict[str, JsonValue]: + return { + "model": _MODEL, + "system": "Keep this context", + "tools": [{"name": "lookup", "input_schema": {"type": "object"}}], + "messages": [{"role": "user", "content": [ + {"type": "text", "text": "A cacheable prefix", "cache_control": {"type": "ephemeral"}} + ]}], + } + + +@pytest.mark.parametrize("version", [None, "2099-01-01", DEFAULT_ANTHROPIC_API_VERSION]) +@pytest.mark.asyncio +async def test_observer_records_only_version_supported_by_token_counter(version: str | None) -> None: + cache: Final = DualCache() + observer: Final = PromptCacheObserver(InternalUsageCache(dual_cache=cache), clock=lambda: 1010.0) + body: Final = _body() + prefix: Final = parse_prompt(body) + assert prefix is not None + headers: Final = {"x-api-key": _KEY, **({"anthropic-version": version} if version is not None else {})} + wire: Final = httpx.Request("POST", "https://api.anthropic.com/v1/messages", headers=headers, json=body) + response: Final = ModelResponse( + model=_MODEL, + usage=Usage( + prompt_tokens=311, + completion_tokens=2, + total_tokens=313, + prompt_tokens_details=PromptTokensDetailsWrapper( + cached_tokens=100, + cache_creation_tokens=200, + cache_creation_token_details=CacheCreationTokenDetails( + ephemeral_5m_input_tokens=200, ephemeral_1h_input_tokens=0 + ), + ), + ), + ) + await observer.async_log_success_event( + { + "call_type": "anthropic_messages", + "custom_llm_provider": "anthropic", + "httpx_response": httpx.Response(200, request=wire), + "first_api_call_start_time": datetime.fromtimestamp(1000.0), + "standard_logging_object": { + "status": "success", "model_id": _DEPLOYMENT, + "metadata": {"user_api_key_hash": _CALLER}, + }, + }, + response, + datetime.fromtimestamp(1010.0), + datetime.fromtimestamp(1010.0), + ) + default_scope: Final = cache_scope(_CALLER, _DEPLOYMENT, _KEY, _MODEL) + found: Final = await lookup(cache, default_scope, prefix, now=1010.0) + assert (found is not None) == (version == DEFAULT_ANTHROPIC_API_VERSION) + if version != DEFAULT_ANTHROPIC_API_VERSION: + other_scope: Final = cache_scope(_CALLER, _DEPLOYMENT, _KEY, _MODEL, version or "") + assert await lookup(cache, other_scope, prefix, now=1010.0) is None + + +@pytest.mark.parametrize("headers, supported", [ + ({}, True), + ({"Anthropic-Version": DEFAULT_ANTHROPIC_API_VERSION}, True), + ({"anthropic-version": "2099-01-01"}, False), + ({"Anthropic-Beta": ""}, False), + ({"anthropic-beta": "future-feature"}, False), +]) +def test_prediction_header_eligibility(headers: Mapping[str, str], supported: bool) -> None: + assert supported_prediction_headers(headers) is supported + + +@pytest.mark.asyncio +async def test_provider_count_uses_same_version_and_preserves_native_input(monkeypatch: pytest.MonkeyPatch) -> None: + body: Final = _body() + requests: Final[list[httpx.Request]] = [] + + def provider(request: httpx.Request) -> httpx.Response: + requests.append(request) + return httpx.Response(200, json={"input_tokens": 311}) + + client: Final = AsyncHTTPHandler() + await client.client.aclose() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(provider)) + monkeypatch.setattr(count_handler, "get_async_httpx_client", lambda **kwargs: client) + try: + assert await count_prompt_tokens(_MODEL, _KEY, body) == 311 + finally: + await client.client.aclose() + assert len(requests) == 1 + assert requests[0].headers["anthropic-version"] == DEFAULT_ANTHROPIC_API_VERSION + assert requests[0].url == "https://api.anthropic.com/v1/messages/count_tokens" + assert json.loads(requests[0].content) == body + + +@pytest.mark.parametrize("source", ["static", "database"]) +@pytest.mark.asyncio +async def test_environment_credential_matches_native_count_and_observed_scope( + source: str, monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("LIT7658_PROVIDER_KEY", _KEY) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(litellm, "in_memory_llm_clients_cache", LLMClientCache()) + params: Final = { + "model": f"anthropic/{_MODEL}", "api_key": "os.environ/LIT7658_PROVIDER_KEY", + "api_base": "https://api.anthropic.com", + } + router: Final = litellm.Router(model_list=[{ + "model_name": "test-native", "litellm_params": dict(params), "model_info": {"id": _DEPLOYMENT}, + }] if source == "static" else [], num_retries=0) + if source == "database": + monkeypatch.setattr(proxy_server, "llm_router", router) + assert proxy_server.ProxyConfig()._add_deployment([SimpleNamespace( + model_id=_DEPLOYMENT, model_name="test-native", model_info={}, litellm_params=dict(params), + )]) == 1 + deployment: Final = router.get_deployment(_DEPLOYMENT) + assert deployment is not None + target: Final = resolve_prediction_target(deployment.litellm_params) + assert isinstance(target, NativePredictionTarget) + body: Final = _body() + with respx.mock() as upstream: + native: Final = upstream.post("https://api.anthropic.com/v1/messages").respond(200, json={ + "id": "msg_test", "type": "message", "role": "assistant", "model": _MODEL, + "content": [{"type": "text", "text": "Hello"}], "stop_reason": "end_turn", "stop_sequence": None, + "usage": {"input_tokens": 11, "output_tokens": 1, "cache_read_input_tokens": 300}, + }) + counter: Final = upstream.post("https://api.anthropic.com/v1/messages/count_tokens").respond( + 200, json={"input_tokens": 311}, + ) + await router.aanthropic_messages( + model="test-native", max_tokens=1, **{key: value for key, value in body.items() if key != "model"}, + ) + assert await count_prompt_tokens(target.model, target.api_key, body) == 311 + assert native.call_count == counter.call_count == 1 + assert native.calls.last.request.headers["x-api-key"] == counter.calls.last.request.headers["x-api-key"] == _KEY + observed: Final = parse_observed_cache(native.calls.last.request, ModelResponse( + model=_MODEL, usage=Usage( + prompt_tokens=311, completion_tokens=1, total_tokens=312, + prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=300), + ), + ), _CALLER, _DEPLOYMENT) + assert observed is not None + assert observed.scope == cache_scope(_CALLER, _DEPLOYMENT, target.api_key, target.model) + + +@pytest.mark.parametrize("inline_key", [None, _KEY]) +@pytest.mark.asyncio +async def test_named_credential_is_explicitly_unsupported_before_count( + inline_key: str | None, monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setattr(litellm, "credential_list", [CredentialItem( + credential_name="test-named", credential_info={}, credential_values={"api_key": "test-named-provider-key"}, + )]) + deployment: Final = Deployment( + model_name="test-native", + litellm_params=LiteLLM_Params( + model=f"anthropic/{_MODEL}", api_key=inline_key, litellm_credential_name="test-named", + ), + model_info=ModelInfo(id=_DEPLOYMENT), + ) + body: Final = _body() + prefix: Final = parse_prompt(body) + assert prefix is not None + + async def count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + pytest.fail("Unsupported named credentials must not reach provider counting") + + arm: Final = await predict_arm(deployment, body, prefix, _CALLER, DualCache(), count) + assert arm.cache_state == "unknown" + assert arm.reason == "unsupported_deployment_configuration" + assert arm.estimate is None and arm.cold is None and arm.warm is None diff --git a/tests/test_litellm/proxy/auth/test_auth_utils.py b/tests/test_litellm/proxy/auth/test_auth_utils.py index cdf1f897707..bd6a14cad21 100644 --- a/tests/test_litellm/proxy/auth/test_auth_utils.py +++ b/tests/test_litellm/proxy/auth/test_auth_utils.py @@ -433,6 +433,232 @@ def test_get_model_from_request_no_request_extracts_model(): ) +def _cache_prediction_router(): + from litellm.router import Router + + return Router(model_list=[ + { + "model_name": group, + "litellm_params": {"model": "anthropic/claude-sonnet-5", "api_key": "test-provider-key"}, + "model_info": {"id": deployment_id, "team_id": team_id}, + } + for group, deployment_id, team_id in ( + ("current-group", "current-id", None), ("candidate-group", "candidate-id", None), + ("own-group", "own-id", "prediction-team"), ("foreign-group", "foreign-id", "foreign-team"), + ) + ]) + + +@pytest.mark.parametrize("candidate,team_id,expected", [ + ("candidate-id", None, ["current-group", "candidate-group"]), + ("current-id", None, "current-group"), + ("missing-id", None, None), + ("candidate-group", None, None), + ("own-id", None, None), + ("own-id", "prediction-team", ["current-group", "own-group"]), + ("foreign-id", "prediction-team", None), +]) +def test_cache_prediction_auth_resolves_only_exact_deployment_ids(candidate, team_id, expected): + assert get_model_from_request( + request_data={ + "current_deployment_id": "current-id", "candidate_deployment_id": candidate, + "request": {"model": "caller-controlled-provider-model"}, + }, + route="/cost/predict-cache", + llm_router=_cache_prediction_router(), + team_id=team_id, + ) == expected + + +def _cache_prediction_auth_app( + monkeypatch, allowed_routes, user_models, metadata=None, *, team_id=None, key_models=None, team_models=None +): + import importlib + from unittest.mock import AsyncMock + + from fastapi import FastAPI + + import litellm.proxy.proxy_server as proxy_server + from litellm.caching.dual_cache import DualCache + from litellm.proxy._types import LiteLLM_TeamTableCachedObj, LiteLLM_UserTable, LitellmUserRoles, ProxyException + from litellm.proxy.auth import auth_checks + from litellm.proxy.hooks.parallel_request_limiter_v3 import _PROXY_MaxParallelRequestsHandler_v3 + from litellm.proxy.management_endpoints import prompt_cache_prediction as endpoint + from litellm.proxy.utils import InternalUsageCache, ProxyLogging + + auth = importlib.import_module("litellm.proxy.auth.user_api_key_auth") + router = _cache_prediction_router() + allowed_models = ["current-group", "candidate-group", "own-group"] + token = UserAPIKeyAuth( + api_key="test-proxy-key-hash", user_id="prediction-user", user_role=LitellmUserRoles.INTERNAL_USER, + models=allowed_models if key_models is None else key_models, team_id=team_id, + team_models=allowed_models if team_models is None else team_models, + allowed_routes=allowed_routes, metadata=metadata or {}, + ) + user = LiteLLM_UserTable( + user_id=token.user_id, user_role=LitellmUserRoles.INTERNAL_USER.value, models=user_models, + ) + async def authenticate(request, request_data, **_headers): + await auth._enforce_key_and_fallback_model_access( + valid_token=token, request_data=request_data, route=request.url.path, request=request, + llm_model_list=router.get_model_list(), llm_router=router, + ) + return token + + monkeypatch.setattr(auth, "_user_api_key_auth_builder", authenticate) + monkeypatch.setattr(auth, "get_user_object", AsyncMock(return_value=user)) + team = LiteLLM_TeamTableCachedObj(team_id=team_id, models=token.team_models) if team_id else None + monkeypatch.setattr(auth, "get_team_object", AsyncMock(return_value=team)) + monkeypatch.setattr(auth_checks, "get_team_object", AsyncMock(return_value=team)) + monkeypatch.setattr(auth_checks, "get_team_membership", AsyncMock(return_value=None)) + monkeypatch.setattr(auth, "get_global_proxy_spend", AsyncMock(return_value=0)) + monkeypatch.setattr(proxy_server, "master_key", "test-master-key") + monkeypatch.setattr(proxy_server, "user_custom_auth", None) + monkeypatch.setattr(proxy_server, "general_settings", {}) + monkeypatch.setattr(proxy_server, "llm_router", router) + monkeypatch.setattr(proxy_server, "llm_model_list", router.get_model_list()) + monkeypatch.setattr(proxy_server, "prisma_client", None) + monkeypatch.setattr(proxy_server, "user_api_key_cache", DualCache()) + logging = ProxyLogging(user_api_key_cache=DualCache()) + logging.proxy_hook_mapping["parallel_request_limiter"] = _PROXY_MaxParallelRequestsHandler_v3( + InternalUsageCache(dual_cache=DualCache()) + ) + monkeypatch.setattr(proxy_server, "proxy_logging_obj", logging) + counts = AsyncMock(return_value=6_000) + monkeypatch.setattr(endpoint, "count_prompt_tokens", counts) + app = FastAPI() + app.include_router(endpoint.router) + app.add_exception_handler(ProxyException, proxy_server.openai_exception_handler) + return app, counts + + +def _cache_prediction_payload(candidate="candidate-id", current="current-id"): + return { + "current_deployment_id": current, "candidate_deployment_id": candidate, + "request": {"messages": [{"role": "user", "content": [{ + "type": "text", "text": "Stable cached context", + "cache_control": {"type": "ephemeral"}, + }]}]}, + } + + +@pytest.mark.asyncio +@pytest.mark.parametrize("allowed_routes,user_models,candidate,status_code", [ + (["/chat/completions"], ["current-group", "candidate-group"], "candidate-id", 403), + (["/cost/predict-cache"], ["current-group"], "candidate-id", 403), + (["/cost/*"], ["current-group", "candidate-group"], "candidate-id", 200), + (["/cost/predict-cache"], ["current-group"], "current-id", 200), + (["/cost/predict-cache"], ["current-group"], "missing-id", 404), +]) +async def test_cache_prediction_authorizes_route_and_personal_models_before_provider_counts( + monkeypatch, allowed_routes, user_models, candidate, status_code +): + import httpx + + app, counts = _cache_prediction_auth_app(monkeypatch, allowed_routes, user_models) + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + response = await client.post("/cost/predict-cache", json=_cache_prediction_payload(candidate)) + + assert response.status_code == status_code, response.text + if status_code == 200: + assert counts.await_count == (2 if candidate == "current-id" else 4) + else: + assert counts.await_count == 0 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("arm", ["current_deployment_id", "candidate_deployment_id"]) +@pytest.mark.parametrize("team_id,key_models,user_models,team_models", [ + (None, ["*"], ["*"], None), + (None, ["current-group", "candidate-group"], ["*"], None), + (None, ["*"], ["current-group", "candidate-group"], None), + ("prediction-team", ["*"], ["*"], ["current-group", "candidate-group"]), +]) +async def test_cache_prediction_hides_foreign_and_missing_ids_before_model_authorization( + monkeypatch, arm, team_id, key_models, user_models, team_models +): + import httpx + + app, counts = _cache_prediction_auth_app( + monkeypatch, ["/cost/predict-cache"], user_models, + team_id=team_id, key_models=key_models, team_models=team_models, + ) + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + missing = await client.post("/cost/predict-cache", json={**_cache_prediction_payload(), arm: "missing-id"}) + foreign = await client.post("/cost/predict-cache", json={**_cache_prediction_payload(), arm: "foreign-id"}) + + assert missing.status_code == foreign.status_code == 404, foreign.text + assert missing.json() == foreign.json() == {"detail": "Deployment not found"} + assert counts.await_count == 0 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("arm", ["current_deployment_id", "candidate_deployment_id"]) +@pytest.mark.parametrize("key_models,team_models,status_code", [ + (["*"], ["*"], 200), + (["current-group", "candidate-group"], ["*"], 403), + (["*"], ["current-group", "candidate-group"], 403), +]) +async def test_cache_prediction_checks_each_visible_team_deployment_model( + monkeypatch, arm, key_models, team_models, status_code +): + import httpx + + app, counts = _cache_prediction_auth_app( + monkeypatch, ["/cost/predict-cache"], ["*"], + team_id="prediction-team", key_models=key_models, team_models=team_models, + ) + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + response = await client.post("/cost/predict-cache", json={**_cache_prediction_payload(), arm: "own-id"}) + + assert response.status_code == status_code, response.text + assert counts.await_count == (4 if status_code == 200 else 0) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("arm", ["current_deployment_id", "candidate_deployment_id"]) +async def test_cache_prediction_checks_each_visible_personal_deployment_model(monkeypatch, arm): + import httpx + + app, counts = _cache_prediction_auth_app(monkeypatch, ["/cost/predict-cache"], ["current-group"]) + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + response = await client.post( + "/cost/predict-cache", json={**_cache_prediction_payload(candidate="current-id"), arm: "candidate-id"} + ) + + assert response.status_code == 403, response.text + assert response.json()["error"]["type"] == "user_model_access_denied" + assert counts.await_count == 0 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("header_tag,key_tags,limit,status_code,provider_calls", [ + ("limited", [], 1, 429, 1), + (None, ["limited"], 1, 429, 1), + ("limited", ["limited"], 4, 200, 4), + ("unlimited", [], 1, 200, 4), +]) +async def test_cache_prediction_preserves_authenticated_header_and_key_tag_rpm( + monkeypatch, header_tag, key_tags, limit, status_code, provider_calls +): + import httpx + + app, counts = _cache_prediction_auth_app( + monkeypatch, ["/cost/predict-cache"], ["current-group", "candidate-group"], + metadata={"tag_rpm_limit": {"limited": limit}, "tags": key_tags}, + ) + headers = {"x-litellm-tags": header_tag} if header_tag else {} + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + response = await client.post("/cost/predict-cache", json=_cache_prediction_payload(), headers=headers) + assert response.status_code == status_code, response.text + assert counts.await_count == provider_calls + if limit == 4: + exhausted = await client.post("/cost/predict-cache", json=_cache_prediction_payload(), headers=headers) + assert exhausted.status_code == 429, exhausted.text + assert counts.await_count == 4 + assert all("metadata" not in call.args[2] for call in counts.await_args_list) + + def test_get_model_from_request_supports_google_model_names_with_slashes(): assert ( get_model_from_request( diff --git a/tests/test_litellm/proxy/common_utils/test_prompt_cache_pricing.py b/tests/test_litellm/proxy/common_utils/test_prompt_cache_pricing.py new file mode 100644 index 00000000000..994684a6005 --- /dev/null +++ b/tests/test_litellm/proxy/common_utils/test_prompt_cache_pricing.py @@ -0,0 +1,105 @@ +from typing import Final + +import pytest + +import litellm +from litellm.proxy.common_utils.prompt_cache_pricing import price_cache_tokens +from litellm.types.management_endpoints.prompt_cache_prediction import CacheTokenBuckets + + +@pytest.mark.parametrize( + ("model", "expected"), + [("anthropic/claude-sonnet-4-5", 1.26), ("anthropic/claude-sonnet-4-6", 0.63)], +) +def test_prices_all_cache_buckets_at_total_context_tier(model: str, expected: float) -> None: + tokens: Final = CacheTokenBuckets( + uncached_input_tokens=100_000, + cache_read_input_tokens=50_000, + cache_creation_5m_input_tokens=20_000, + cache_creation_1h_input_tokens=40_000, + ) + assert price_cache_tokens(model, "unconfigured-deployment", tokens) == pytest.approx(expected) + + +@pytest.mark.parametrize(("total", "expected"), [(200_000, 0.387), (200_001, 0.774006)]) +def test_long_context_tier_starts_above_threshold(total: int, expected: float) -> None: + tokens: Final = CacheTokenBuckets( + uncached_input_tokens=total - 100_000, + cache_creation_1h_input_tokens=10_000, + cache_read_input_tokens=90_000, + ) + actual: Final = price_cache_tokens("anthropic/claude-sonnet-4-5", "unconfigured-deployment", tokens) + assert actual == pytest.approx(expected) + + +def test_deployment_tariff_wins_without_proxy_discounts_or_margins(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setattr(litellm, "model_cost", litellm.model_cost.copy()) + litellm.Router( + model_list=[ + { + "model_name": "cache-pricing-test", + "litellm_params": { + "model": "anthropic/claude-sonnet-4-6", + "api_key": "test-only", + "input_cost_per_token": 0.00001, + "output_cost_per_token": 0.00002, + "cache_read_input_token_cost": 0.000001, + "cache_creation_input_token_cost": 0.0000125, + "cache_creation_input_token_cost_above_1hr": 0.00002, + }, + "model_info": {"id": "cache-pricing-test-a"}, + } + ] + ) + monkeypatch.setattr(litellm, "cost_discount_config", {"anthropic": 0.5}) + monkeypatch.setattr(litellm, "cost_margin_config", {"global": {"percentage": 0.3, "fixed_amount": 1.0}}) + tokens: Final = CacheTokenBuckets( + uncached_input_tokens=3_000, + cache_read_input_tokens=4_000, + cache_creation_5m_input_tokens=1_000, + cache_creation_1h_input_tokens=2_000, + ) + assert price_cache_tokens("anthropic/claude-sonnet-4-6", "cache-pricing-test-a", tokens) == pytest.approx(0.0865) + + +@pytest.mark.parametrize("rate", [None, -1.0, float("nan"), float("inf"), "0.00001", True]) +def test_unknown_for_absent_or_invalid_active_cache_rate(monkeypatch: pytest.MonkeyPatch, rate: object) -> None: + monkeypatch.setitem( + litellm.model_cost, + "cache-pricing-invalid", + { + "litellm_provider": "anthropic", + "mode": "chat", + "input_cost_per_token": 0.00001, + "output_cost_per_token": 0.00002, + "cache_creation_input_token_cost_above_1hr": rate, + }, + ) + tokens: Final = CacheTokenBuckets(cache_creation_1h_input_tokens=4_000) + assert price_cache_tokens("anthropic/claude-sonnet-4-6", "cache-pricing-invalid", tokens) is None + + +def test_missing_input_price_is_unknown_even_when_get_model_info_defaults_to_zero( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setitem(litellm.model_cost, "cache-pricing-missing", {"litellm_provider": "anthropic", "mode": "chat"}) + tokens: Final = CacheTokenBuckets(uncached_input_tokens=4_000) + assert price_cache_tokens("cache-pricing-missing", "unconfigured-deployment", tokens) is None + + +def test_explicit_free_pricing_is_not_unknown(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setitem( + litellm.model_cost, + "cache-pricing-free", + { + "litellm_provider": "anthropic", + "mode": "chat", + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + "cache_read_input_token_cost": 0.0, + "cache_creation_input_token_cost": 0.0, + "cache_creation_input_token_cost_above_1hr": 0.0, + }, + ) + tokens: Final = CacheTokenBuckets(uncached_input_tokens=100, cache_read_input_tokens=5_000) + assert price_cache_tokens("anthropic/claude-sonnet-4-6", "cache-pricing-free", tokens) == 0.0 diff --git a/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py b/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py index 10c0bb88a82..48f980086fd 100644 --- a/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py +++ b/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py @@ -6284,3 +6284,248 @@ async def test_an_open_circuit_breaker_reads_the_sliding_window_locally_without_ assert isinstance(values, list) assert [record.getMessage() for record in caplog.records if record.levelno >= logging.WARNING] == [] assert any("circuit breaker is open" in record.getMessage() for record in caplog.records) + + +@pytest.mark.parametrize( + "limits, request_data, counter_scope", + [ + ({"rpm_limit": 1}, {}, "api_key"), + ({"user_id": "u", "user_rpm_limit": 1}, {}, "user"), + ({"team_id": "t", "team_rpm_limit": 1}, {}, "team"), + ( + {"team_id": "t", "user_id": "u", "team_member_rpm_limit": 1}, + {}, + "team_member", + ), + ({"end_user_id": "e", "end_user_rpm_limit": 1}, {}, "end_user"), + ( + {"metadata": {"model_rpm_limit": {"test-model": 1}}}, + {}, + "model_per_key", + ), + ( + {"metadata": {"tag_rpm_limit": {"test-tag": 1}}}, + {"metadata": {"tags": ["test-tag"]}}, + "tag_per_key", + ), + ( + { + "team_id": "t", + "metadata": {"model_rpm_limit": {"test-model": 100}}, + "team_metadata": {"model_rpm_limit": {"test-model": 1}}, + }, + {}, + "model_per_team", + ), + ( + {"project_id": "p", "project_metadata": {"model_rpm_limit": {"test-model": 1}}}, + {}, + "model_per_project", + ), + ({"org_id": "o", "organization_rpm_limit": 1}, {}, "organization"), + ( + {"org_id": "o", "organization_metadata": {"model_rpm_limit": {"test-model": 1}}}, + {}, + "model_per_organization", + ), + ], +) +@pytest.mark.parametrize("request_kind", ["count", "generation"]) +@pytest.mark.asyncio +async def test_request_capacity_enforces_shared_rpm_scopes( + limits, request_data, counter_scope, request_kind +): + cache = DualCache() + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(cache)) + auth = UserAPIKeyAuth(api_key=hash_token("sk-count-rpm"), **limits) + async def request(): + if request_kind == "generation": + await handler.async_pre_call_hook( + user_api_key_dict=auth, + cache=cache, + data={**request_data, "model": "test-model"}, + call_type="acompletion", + ) + return + async with handler.request_capacity(auth, "test-model", request_data=request_data): + pass + + await request() + with pytest.raises(HTTPException) as exc: + await request() + assert exc.value.status_code == 429 + assert counter_scope in str(exc.value.detail) + + +@pytest.mark.asyncio +async def test_request_capacity_keeps_dynamic_rpm_policy(monkeypatch): + import litellm.proxy.proxy_server as proxy_server + + router = Router(model_list=[{ + "model_name": "test-model", + "litellm_params": {"model": "openai/gpt-test", "api_key": "test-key"}, + "model_info": {"id": "test-deployment"}, + }]) + monkeypatch.setattr(proxy_server, "llm_router", router) + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(DualCache())) + auth = UserAPIKeyAuth( + api_key=hash_token("sk-count-dynamic"), + rpm_limit=1, + metadata={"rpm_limit_type": "dynamic"}, + ) + for _ in range(2): + async with handler.request_capacity(auth, "test-model"): + pass + router.cache.set_cache("test-deployment:fails", 100, ttl=60, local_only=True) + async with handler.request_capacity(auth, "test-model"): + pass + with pytest.raises(HTTPException) as exc: + async with handler.request_capacity(auth, "test-model"): + pytest.fail("dynamic RPM must enforce after deployment failures") + assert exc.value.status_code == 429 + + +@pytest.mark.asyncio +async def test_request_capacity_skips_tokens_and_preserves_parent_stash(): + cache = DualCache() + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(cache)) + auth = UserAPIKeyAuth( + api_key=hash_token("sk-count-tpm"), + rpm_limit=5, + tpm_limit=1, + max_parallel_requests=1, + project_id="p", + project_metadata={ + "model_tpm_limit": {"test-model": 1}, + "model_itpm_limit": {"test-model": 1}, + "model_otpm_limit": {"test-model": 1}, + }, + ) + token_scopes = ( + ("api_key", auth.api_key), + ("model_per_project", "p:test-model"), + ("model_per_project_itpm", "p:test-model"), + ("model_per_project_otpm", "p:test-model"), + ) + for scope, value in token_scopes: + token_key = handler.create_rate_limit_keys(scope, value, "tokens") + await cache.async_set_cache(token_key, 100, ttl=60) + await cache.async_set_cache(f"{{{scope}:{value}}}:window", int(time.time()), ttl=60) + parent = get_or_create_request_stash() + parent.reserved_tokens = 123 + parent.parallel_slot = ParallelSlotAcquisition(slot_id="parent", counter_keys=["parent-gauge"]) + for _ in range(2): + async with handler.request_capacity(auth, "test-model"): + assert get_request_stash() is parent + assert parent.parallel_slot["slot_id"] == "parent" + assert parent.reserved_tokens == 123 + for scope, value in token_scopes: + assert await cache.async_get_cache(handler.create_rate_limit_keys(scope, value, "tokens")) == 100 + + +@pytest.mark.parametrize("exit_mode", ["success", "failure", "cancel"]) +@pytest.mark.asyncio +async def test_request_capacity_releases_exact_parallel_slot(exit_mode): + cache = DualCache() + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(cache)) + auth = UserAPIKeyAuth(api_key=hash_token("sk-count-parallel"), max_parallel_requests=1) + entered = asyncio.Event() + finish = asyncio.Event() + + async def provider(): + async with handler.request_capacity(auth, "test-model"): + entered.set() + await finish.wait() + if exit_mode == "failure": + raise RuntimeError("provider failed") + + task = asyncio.create_task(provider()) + await asyncio.wait_for(entered.wait(), timeout=2) + try: + for _ in range(2): + with pytest.raises(HTTPException) as exc: + async with handler.request_capacity(auth, "test-model"): + pytest.fail("rejected request freed the occupied slot") + assert exc.value.status_code == 429 + finally: + if exit_mode == "cancel": + task.cancel() + else: + finish.set() + if exit_mode == "success": + await task + else: + with pytest.raises(asyncio.CancelledError if exit_mode == "cancel" else RuntimeError): + await task + async with handler.request_capacity(auth, "test-model"): + pass + + +class _DelayedCapacityUsageCache: + def __init__(self): + self.delegate = InternalUsageCache(DualCache()) + self.dual_cache = self.delegate.dual_cache + self.acquired = asyncio.Event() + self.finish_admission = asyncio.Event() + self.releasing = asyncio.Event() + self.finish_release = asyncio.Event() + + async def async_get_cache(self, *args, **kwargs): + return await self.delegate.async_get_cache(*args, **kwargs) + + async def async_batch_get_cache(self, *args, **kwargs): + return await self.delegate.async_batch_get_cache(*args, **kwargs) + + async def async_set_cache(self, key, value, **kwargs): + await self.delegate.async_set_cache(key=key, value=value, **kwargs) + if not key.endswith(":max_parallel_requests"): + return + if value: + self.acquired.set() + await self.finish_admission.wait() + else: + self.releasing.set() + await self.finish_release.wait() + + +@pytest.mark.asyncio +async def test_request_capacity_finishes_admission_and_release_despite_repeated_cancel(): + cache = _DelayedCapacityUsageCache() + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=cache) + auth = UserAPIKeyAuth(api_key=hash_token("sk-count-cancel-admission"), max_parallel_requests=1) + + async def provider(): + async with handler.request_capacity(auth, "test-model"): + pytest.fail("cancelled admission entered provider body") + + task = asyncio.create_task(provider()) + await asyncio.wait_for(cache.acquired.wait(), timeout=2) + task.cancel() + await asyncio.sleep(0) + cache.finish_admission.set() + await asyncio.wait_for(cache.releasing.wait(), timeout=2) + task.cancel() + await asyncio.sleep(0) + task.cancel() + await asyncio.sleep(0) + assert not task.done() + cache.finish_release.set() + with pytest.raises(asyncio.CancelledError): + await asyncio.wait_for(task, timeout=2) + async with handler.request_capacity(auth, "test-model"): + pass + + +@pytest.mark.asyncio +async def test_request_capacity_rejection_keeps_existing_redis_mirror(): + cache = DualCache() + handler = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(cache)) + auth = UserAPIKeyAuth(api_key=hash_token("sk-count-mirror"), max_parallel_requests=1) + counter_key = handler.create_rate_limit_keys("api_key", auth.api_key, "max_parallel_requests") + await cache.async_set_cache(counter_key, 1, ttl=60, local_only=True) + for _ in range(2): + with pytest.raises(HTTPException) as exc: + async with handler.request_capacity(auth, "test-model"): + pytest.fail("rejection released another request's mirrored slot") + assert exc.value.status_code == 429 + assert await cache.async_get_cache(counter_key, local_only=True) == 1 diff --git a/tests/test_litellm/proxy/hooks/test_prompt_cache_observer.py b/tests/test_litellm/proxy/hooks/test_prompt_cache_observer.py new file mode 100644 index 00000000000..82af3e9a6ef --- /dev/null +++ b/tests/test_litellm/proxy/hooks/test_prompt_cache_observer.py @@ -0,0 +1,300 @@ +import asyncio +import json +import time +from datetime import datetime + +import httpx +import pytest + +import litellm +from litellm.caching.dual_cache import DualCache +from litellm.llms.anthropic.chat.transformation import AnthropicConfig +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler +from litellm.llms.anthropic.prompt_cache_prediction import cache_scope, parse_prompt +from litellm.proxy.hooks.prompt_cache_prediction import ( + PromptCacheObserver, + lookup, +) +from litellm.proxy.utils import InternalUsageCache +from litellm.types.utils import ModelResponse + +MODEL = "claude-sonnet-5" +CALLER = "a" * 64 +DEPLOYMENT = "native-deployment" +KEY = "test-provider-key" + + +def body(ttl="5m", texts=("private cache prefix",)): + return { + "model": MODEL, + "max_tokens": 2, + "system": "private system instructions", + "tools": [{"name": "lookup", "input_schema": {"type": "object"}}], + "messages": [{"role": "user", "content": [ + {"type": "text", "text": text, **( + {"cache_control": {"type": "ephemeral", "ttl": ttl}} + if index == len(texts) - 1 else {} + )} + for index, text in enumerate(texts) + ]}], + } + + +def usage(ttl="5m", read=100, write=200): + return { + "input_tokens": 11, + "output_tokens": 2, + "cache_read_input_tokens": read, + "cache_creation_input_tokens": write, + "cache_creation": { + "ephemeral_5m_input_tokens": write if ttl == "5m" else 0, + "ephemeral_1h_input_tokens": write if ttl == "1h" else 0, + }, + } + + +def event(request_body, started=1000.0, headers=None, **overrides): + request = httpx.Request( + "POST", "https://api.anthropic.com/v1/messages", json=request_body, + headers={"x-api-key": KEY, "anthropic-version": "2023-06-01", **(headers or {})}, + ) + return { + "call_type": "anthropic_messages", + "custom_llm_provider": "anthropic", + "cache_hit": False, + "httpx_response": httpx.Response(200, request=request), + "first_api_call_start_time": datetime.fromtimestamp(started), + "standard_logging_object": { + "status": "success", "model_id": DEPLOYMENT, + "metadata": {"user_api_key_hash": CALLER}, + }, + **overrides, + } + + +async def observe(cache, request_body=None, native_usage=None, now=1010.0, **overrides): + observer = PromptCacheObserver(InternalUsageCache(dual_cache=cache), clock=lambda: now) + response = ModelResponse( + model=MODEL, + usage=AnthropicConfig().calculate_usage(native_usage or usage(), reasoning_content=None), + ) + await observer.async_log_success_event( + event(request_body or body(), **overrides), response, + datetime.fromtimestamp(now), datetime.fromtimestamp(now), + ) + + +def scope(**overrides): + return cache_scope(**{ + "caller_key_hash": CALLER, "deployment_id": DEPLOYMENT, + "provider_key": KEY, "model": MODEL, **overrides, + }) + + +@pytest.mark.parametrize("ttl,expires", [("5m", 1300), ("1h", 4600)]) +@pytest.mark.asyncio +async def test_observed_cache_count_and_request_start_expiry_survive_as_stale(ttl, expires): + cache = DualCache() + request_body = body(ttl=ttl) + await observe(cache, request_body, usage(ttl=ttl)) + prefix = parse_prompt(request_body) + observed = await lookup(cache, scope(), prefix, now=1200) + assert observed.cached_tokens == 300 + assert observed.observed_at == 1010 + assert observed.expires_at == expires + assert await lookup(cache, scope(), prefix, now=expires) == observed + saved = json.dumps(cache.in_memory_cache.cache_dict) + assert "private cache prefix" not in saved + assert "private system instructions" not in saved + assert KEY not in saved + assert CALLER not in saved + + +@pytest.mark.parametrize("changed", [ + {"caller_key_hash": "b" * 64}, {"deployment_id": "other"}, + {"provider_key": "rotated"}, {"model": "claude-opus-5"}, + {"anthropic_version": "different"}, +]) +@pytest.mark.asyncio +async def test_cache_evidence_is_isolated_by_every_scope_dimension(changed): + cache = DualCache() + await observe(cache) + assert await lookup(cache, scope(**changed), parse_prompt(body()), now=1010) is None + + +@pytest.mark.asyncio +async def test_append_only_prefix_finds_prior_evidence_but_edit_or_context_change_does_not(): + cache = DualCache() + await observe(cache) + extended = parse_prompt(body(texts=("private cache prefix", "new turn"))) + prior = await lookup(cache, scope(), extended, now=1010) + assert prior.cached_tokens == 300 + assert prior.fingerprint != extended.fingerprint + for changed in ( + body(texts=("edited prefix", "new turn")), + {**body(), "system": "different system"}, + {**body(), "tools": [{"name": "other", "input_schema": {"type": "object"}}]}, + body(ttl="1h"), + ): + assert await lookup(cache, scope(), parse_prompt(changed), now=1010) is None + outside_lookback = parse_prompt(body(texts=("private cache prefix", *[str(i) for i in range(20)]))) + assert await lookup(cache, scope(), outside_lookback, now=1010) is None + + +@pytest.mark.parametrize("change", [ + {"thinking": {"type": "enabled", "budget_tokens": 1024}}, + {"tool_choice": {"type": "auto"}}, + {"cache_control": {"type": "ephemeral"}}, + {"tools": [{"type": "web_search_20250305", "name": "web_search"}]}, + {"system": [{"type": "text", "text": "system", "cache_control": {"type": "ephemeral"}}]}, + {"messages": [{"role": "user", "content": [{"type": "image", "source": {}}]}]}, + {"messages": [{"role": "user", "content": "no breakpoint"}]}, +]) +def test_unsupported_or_ambiguous_shapes_have_no_cache_identity(change): + assert parse_prompt({**body(), **change}) is None + duplicate = body() + duplicate["messages"][0]["content"].append(duplicate["messages"][0]["content"][0]) + assert parse_prompt(duplicate) is None + + +@pytest.mark.parametrize("overrides", [ + {"cache_hit": True}, {"call_type": "completion"}, + {"custom_llm_provider": "bedrock"}, {"stream": True}, + {"headers": {"anthropic-beta": "unverified-feature"}}, + {"headers": {"x-custom-header": "unverified"}}, + {"standard_logging_object": {"status": "success", "model_id": DEPLOYMENT, "metadata": {}}}, +]) +@pytest.mark.asyncio +async def test_unverified_source_never_creates_observations(overrides): + cache = DualCache() + await observe(cache, **overrides) + assert await lookup(cache, scope(), parse_prompt(body()), now=1010) is None + + +@pytest.mark.parametrize("native_usage", [ + usage(write=0), + {**usage(), "cache_creation": None}, + {**usage(), "cache_creation": {"ephemeral_5m_input_tokens": 199, "ephemeral_1h_input_tokens": 0}}, + usage(ttl="1h"), + {**usage(), "cache_creation_input_tokens": -200}, +]) +@pytest.mark.asyncio +async def test_missing_or_contradictory_telemetry_cannot_create_observations(native_usage): + cache = DualCache() + await observe(cache, native_usage=native_usage) + assert await lookup(cache, scope(), parse_prompt(body()), now=1010) is None + + +@pytest.mark.asyncio +async def test_pure_read_refresh_requires_prior_matching_evidence(): + cache = DualCache() + await observe(cache, native_usage=usage(read=300, write=0)) + assert await lookup(cache, scope(), parse_prompt(body()), now=1010) is None + await observe(cache) + await observe(cache, native_usage=usage(read=300, write=0), started=1100, now=1110) + assert (await lookup(cache, scope(), parse_prompt(body()), now=1110)).expires_at == 1400 + + +class RecordingObserver(PromptCacheObserver): + def __init__(self, cache): + super().__init__(InternalUsageCache(dual_cache=cache)) + self.finished = asyncio.Event() + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time): + await super().async_log_success_event(kwargs, response_obj, start_time, end_time) + self.finished.set() + + +def native_response(): + return { + "id": "msg_prediction", "type": "message", "role": "assistant", "model": MODEL, + "content": [{"type": "text", "text": "ok"}], "stop_reason": "end_turn", + "stop_sequence": None, "usage": usage(ttl="1h"), + } + + +def stream_response(completed, provider_error=False): + response = native_response() + events = [ + {"type": "message_start", "message": {**response, "content": [], "stop_reason": None}}, + {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}, + {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "ok"}}, + {"type": "content_block_stop", "index": 0}, + {"type": "message_delta", "delta": {"stop_reason": "end_turn"}, "usage": {"output_tokens": 2}}, + ] + if completed: + events.append({"type": "message_stop"}) + if provider_error: + events.append({"type": "error", "error": {"type": "overloaded_error", "message": "temporary failure"}}) + return "".join(f"event: {item['type']}\ndata: {json.dumps(item)}\n\n" for item in events) + + +class TransportChunks(httpx.AsyncByteStream): + def __init__(self, payload, chunk_size, fragment_error_only=False): + self.payload = payload.encode() + self.chunk_size = chunk_size or len(self.payload) + self.prefix_length = self.payload.index(b"event: error") if fragment_error_only else 0 + + async def __aiter__(self): + if self.prefix_length: + yield self.payload[:self.prefix_length] + for offset in range(self.prefix_length, len(self.payload), self.chunk_size): + yield self.payload[offset:offset + self.chunk_size] + + +@pytest.mark.parametrize("stream,completed,provider_error,transport", [ + (False, True, False, "whole"), + (True, True, False, "whole"), + (True, False, False, "whole"), + (True, True, True, "whole"), + (True, True, False, "fragmented"), + (True, False, False, "fragmented"), + (True, True, True, "fragmented"), + (True, True, True, "fragmented_error"), + (True, True, False, "unterminated"), +]) +@pytest.mark.asyncio +async def test_native_production_callback_records_only_completed_wire_requests(stream, completed, provider_error, transport): + cache = DualCache() + observer = RecordingObserver(cache) + litellm.logging_callback_manager.add_litellm_callback(observer) + + def provider(request): + if stream: + payload = stream_response(completed, provider_error) + if transport == "unterminated": + payload = payload.removesuffix("\n\n") + return httpx.Response( + 200, request=request, headers={"content-type": "text/event-stream"}, + stream=TransportChunks( + payload, 1 if transport.startswith("fragmented") else None, + fragment_error_only=transport == "fragmented_error", + ), + ) + return httpx.Response(200, request=request, json=native_response()) + + client = AsyncHTTPHandler() + await client.client.aclose() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(provider)) + try: + request_body = body(ttl="1h") + before = time.time() + result = await litellm.anthropic_messages( + **{**request_body, "model": f"anthropic/{MODEL}"}, + api_key=KEY, client=client, stream=stream, model_info={"id": DEPLOYMENT}, + litellm_metadata={"user_api_key_hash": CALLER, "model_info": {"id": DEPLOYMENT}}, + ) + if stream: + async for _ in result: + pass + await asyncio.wait_for(observer.finished.wait(), timeout=5) + found = await lookup(cache, scope(), parse_prompt(request_body)) + if completed and not provider_error and transport != "unterminated": + assert found is not None + assert found.cached_tokens == 300 + assert before + 3600 <= found.expires_at <= time.time() + 3600 + else: + assert found is None + finally: + litellm.logging_callback_manager.remove_callback_from_all_lists(observer) + await client.client.aclose() diff --git a/tests/test_litellm/proxy/management_endpoints/test_prompt_cache_prediction.py b/tests/test_litellm/proxy/management_endpoints/test_prompt_cache_prediction.py new file mode 100644 index 00000000000..0ec277be884 --- /dev/null +++ b/tests/test_litellm/proxy/management_endpoints/test_prompt_cache_prediction.py @@ -0,0 +1,698 @@ +import asyncio +import time +from collections.abc import Iterator, Mapping +from dataclasses import dataclass +from typing import Final, Literal + +import httpx +import pytest +from fastapi import FastAPI, Request +from pydantic import JsonValue + +import litellm +from litellm.caching.caching import DualCache +from litellm.integrations.custom_logger import CustomLogger +from litellm.proxy._types import ProxyException, UserAPIKeyAuth +from litellm.proxy.common_utils.http_parsing_utils import _read_request_body, _safe_set_request_parsed_body +from litellm.proxy.hooks.parallel_request_limiter_v3 import _PROXY_MaxParallelRequestsHandler_v3 +from litellm.llms.anthropic.prompt_cache_prediction import PromptPrefix, cache_scope, parse_prompt +from litellm.proxy.hooks.prompt_cache_prediction import ( + CacheObservation, + _cache_key, +) +from litellm.proxy.management_endpoints import prompt_cache_prediction as endpoint +from litellm.proxy.utils import InternalUsageCache +from litellm.types.management_endpoints.prompt_cache_prediction import CachePredictionResponse +from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo + + +_PROVIDER_KEY: Final = "cache-prediction-test-provider-key" +_CALLER: Final = "cache-prediction-test-caller-hash" + + +@pytest.fixture(autouse=True) +def anthropic_endpoint_environment(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.delenv("ANTHROPIC_API_BASE", raising=False) + monkeypatch.delenv("ANTHROPIC_BASE_URL", raising=False) + + +def _body(ttl: str = "5m", *, extended: bool = False) -> dict[str, JsonValue]: + blocks: Final[list[JsonValue]] = [ + {"type": "text", "text": "Stable context"}, + *([{"type": "text", "text": "Appended context"}] if extended else []), + ] + return { + "max_tokens": 10, + "system": "Follow the project conventions", + "messages": [ + { + "role": "user", + "content": [ + *blocks[:-1], + {**blocks[-1], "cache_control": {"type": "ephemeral", "ttl": ttl}}, + {"type": "text", "text": "Follow-up question"}, + ], + } + ], + } + + +def _prefix(body: Mapping[str, JsonValue]) -> PromptPrefix: + prefix: Final = parse_prompt(body) + assert prefix is not None + return prefix + + +def _deployment( + deployment_id: str = "sonnet", + model: str = "claude-sonnet-5", + *, + team_id: str | None = None, + api_base: str | None = None, +) -> Deployment: + return Deployment( + model_name=deployment_id, + litellm_params=LiteLLM_Params(model=f"anthropic/{model}", api_key=_PROVIDER_KEY, api_base=api_base), + model_info=ModelInfo(id=deployment_id, team_id=team_id), + ) + + +@dataclass(frozen=True) +class Counts: + total: int | None = 6_000 + prefix: int | None = 5_000 + + async def __call__(self, model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + assert api_key == _PROVIDER_KEY + assert model.startswith("claude-") + return self.total if "max_tokens" in body else self.prefix + + +async def _observe( + cache: DualCache, + body: Mapping[str, JsonValue], + *, + deployment_id: str = "sonnet", + model: str = "claude-sonnet-5", + cached_tokens: int = 5_000, + expired: bool = False, + caller: str = _CALLER, +) -> None: + prefix: Final = _prefix(body) + now: Final = time.time() + observation: Final = CacheObservation( + fingerprint=prefix.fingerprint, + cached_tokens=cached_tokens, + observed_at=now - 400 if expired else now - 10, + expires_at=now - 100 if expired else now + 290, + ) + scope: Final = cache_scope(caller, deployment_id, _PROVIDER_KEY, model) + await cache.async_set_cache(_cache_key(scope, prefix.fingerprint), observation.model_dump_json(), ttl=3_600) + + +@pytest.mark.asyncio +@pytest.mark.parametrize(("ttl", "cold_cost"), [("5m", 0.0145), ("1h", 0.022)]) +async def test_unobserved_cache_prices_cold_and_warm_bounds(ttl: str, cold_cost: float) -> None: + body: Final = _body(ttl) + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, DualCache(), Counts()) + + assert arm.cache_state == "unknown" + assert arm.reason == "no_compatible_observation" + assert arm.evidence is None + assert arm.estimate is not None and arm.cold is not None and arm.warm is not None + assert arm.estimate.input_cost == pytest.approx(cold_cost) + assert arm.cold.input_cost == pytest.approx(cold_cost) + assert arm.warm.input_cost == pytest.approx(0.003) + assert arm.cold.tokens.uncached_input_tokens == 1_000 + assert arm.cold.tokens.cache_read_input_tokens == 0 + assert arm.cold.tokens.cache_creation_5m_input_tokens == (5_000 if ttl == "5m" else 0) + assert arm.cold.tokens.cache_creation_1h_input_tokens == (5_000 if ttl == "1h" else 0) + assert arm.warm.tokens.cache_read_input_tokens == 5_000 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("cached_tokens", "warm_cost", "cold_cost"), [(5_400, 0.00228, 0.0147), (4_600, 0.00372, 0.0143)] +) +@pytest.mark.parametrize("expired", [False, True]) +async def test_exact_prefix_conserves_total_with_observed_count_in_all_scenarios( + cached_tokens: int, warm_cost: float, cold_cost: float, expired: bool +) -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body, cached_tokens=cached_tokens, expired=expired) + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, cache, Counts()) + + assert arm.cache_state == ("stale" if expired else "warm") + assert arm.evidence is not None + assert arm.estimate is not None and arm.warm is not None and arm.cold is not None + assert arm.warm.tokens.cache_read_input_tokens == cached_tokens + assert arm.warm.tokens.cache_creation_5m_input_tokens == 0 + assert arm.cold.tokens.cache_creation_5m_input_tokens == cached_tokens + assert arm.cold.tokens.cache_read_input_tokens == 0 + for scenario in (arm.estimate, arm.cold, arm.warm): + assert scenario.tokens.total_tokens == 6_000 + assert scenario.tokens.uncached_input_tokens == 6_000 - cached_tokens + assert arm.warm.input_cost == pytest.approx(warm_cost) + assert arm.cold.input_cost == pytest.approx(cold_cost) + assert arm.estimate.input_cost == pytest.approx(cold_cost if expired else warm_cost) + + +@pytest.mark.asyncio +async def test_observed_prefix_larger_than_full_request_returns_unknown() -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body, cached_tokens=6_001) + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, cache, Counts()) + + assert arm.cache_state == "unknown" + assert arm.reason == "inconsistent_prefix_token_count" + assert arm.estimate is None and arm.cold is None and arm.warm is None + + +@pytest.mark.asyncio +@pytest.mark.parametrize(("ttl", "expected"), [("5m", 0.0053), ("1h", 0.0068)]) +async def test_append_only_prefix_reads_old_tokens_and_writes_extension(ttl: str, expected: float) -> None: + cache: Final = DualCache() + await _observe(cache, _body(ttl), cached_tokens=4_000) + body: Final = _body(ttl, extended=True) + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, cache, Counts()) + + assert arm.cache_state == "partial" + assert arm.estimate is not None + assert arm.estimate.tokens.cache_read_input_tokens == 4_000 + assert arm.estimate.tokens.cache_creation_5m_input_tokens == (1_000 if ttl == "5m" else 0) + assert arm.estimate.tokens.cache_creation_1h_input_tokens == (1_000 if ttl == "1h" else 0) + assert arm.estimate.input_cost == pytest.approx(expected) + + +@pytest.mark.asyncio +async def test_expired_observation_estimates_a_cold_rebuild() -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body, expired=True) + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, cache, Counts()) + + assert arm.cache_state == "stale" + assert arm.reason == "observation_expired" + assert arm.evidence is not None and arm.evidence.expires_at < time.time() + assert arm.estimate is not None and arm.cold is not None + assert arm.estimate.tokens.cache_read_input_tokens == 0 + assert arm.estimate.tokens.cache_creation_5m_input_tokens == 5_000 + assert arm.estimate.input_cost == arm.cold.input_cost + + +@pytest.mark.asyncio +async def test_below_model_minimum_prices_all_input_as_uncached() -> None: + body: Final = _body() + arm: Final = await endpoint.predict_arm( + _deployment(), body, _prefix(body), _CALLER, DualCache(), Counts(total=1_500, prefix=1_000) + ) + + assert arm.cache_state == "disabled" + assert arm.reason == "below_cache_minimum" + assert arm.estimate is not None + assert arm.estimate.tokens.uncached_input_tokens == 1_500 + assert arm.estimate.tokens.cache_read_input_tokens == 0 + assert arm.estimate.tokens.cache_creation_5m_input_tokens == 0 + assert arm.estimate.input_cost == pytest.approx(0.003) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("counts", [Counts(total=None), Counts(prefix=None), Counts(total=4_000)]) +async def test_unavailable_or_inconsistent_token_counts_return_null_estimates(counts: Counts) -> None: + body: Final = _body() + arm: Final = await endpoint.predict_arm(_deployment(), body, _prefix(body), _CALLER, DualCache(), counts) + + assert arm.cache_state == "unknown" + assert arm.reason == "token_count_unavailable" + assert arm.estimate is None and arm.cold is None and arm.warm is None + + +@pytest.mark.asyncio +@pytest.mark.parametrize("counts", [Counts(), Counts(total=1_500, prefix=1_000)]) +async def test_missing_prices_return_unknown_and_null_estimates( + monkeypatch: pytest.MonkeyPatch, counts: Counts +) -> None: + monkeypatch.setitem( + litellm.model_cost, + "claude-cache-unpriced-5", + {"litellm_provider": "anthropic", "mode": "chat"}, + ) + body: Final = _body() + arm: Final = await endpoint.predict_arm( + _deployment("cache-prediction-unpriced", "claude-cache-unpriced-5"), + body, + _prefix(body), + _CALLER, + DualCache(), + counts, + ) + + assert arm.cache_state == "unknown" + assert arm.reason == "pricing_unavailable" + assert arm.estimate is None and arm.cold is None and arm.warm is None + + +@pytest.mark.asyncio +async def test_custom_api_base_from_environment_returns_unknown_before_counting( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("ANTHROPIC_API_BASE", "https://custom.invalid") + body: Final = _body() + arm: Final = await endpoint.predict_arm( + _deployment(), body, _prefix(body), _CALLER, DualCache(), _unexpected_count + ) + + assert arm.cache_state == "unknown" + assert arm.reason == "unsupported_provider_endpoint" + assert arm.estimate is None and arm.cold is None and arm.warm is None + + +@pytest.mark.asyncio +async def test_explicit_official_api_base_overrides_custom_environment(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("ANTHROPIC_API_BASE", "https://custom.invalid") + body: Final = _body() + arm: Final = await endpoint.predict_arm( + _deployment(api_base="https://api.anthropic.com"), body, _prefix(body), _CALLER, DualCache(), Counts() + ) + + assert arm.cache_state == "unknown" + assert arm.reason == "no_compatible_observation" + assert arm.estimate is not None + assert arm.estimate.input_cost == pytest.approx(0.0145) + + +@dataclass(frozen=True) +class _ProxyLogging: + internal_usage_cache: InternalUsageCache + parallel_limiter: CustomLogger | None + + def get_proxy_hook(self, hook: str) -> CustomLogger | None: + return self.parallel_limiter if hook == "parallel_request_limiter" else None + + +def _app( + monkeypatch: pytest.MonkeyPatch, + cache: DualCache, + *, + caller: UserAPIKeyAuth | None = None, + current_team: str | None = None, + candidate_team: str | None = None, + counts: endpoint.TokenCounter = Counts(), + limiter: CustomLogger | Literal["default"] | None = "default", +) -> FastAPI: + import litellm.proxy.proxy_server as proxy_server + + model_list: Final = [ + _deployment("opus", "claude-opus-5", team_id=current_team).model_dump(exclude_unset=True), + _deployment("sonnet", team_id=candidate_team).model_dump(exclude_unset=True), + ] + router: Final = litellm.Router(model_list=model_list) + monkeypatch.setattr(proxy_server, "llm_router", router) + monkeypatch.setattr(proxy_server, "llm_model_list", model_list) + monkeypatch.setattr(endpoint, "count_prompt_tokens", counts) + app: Final = FastAPI() + app.include_router(endpoint.router) + app.add_exception_handler(ProxyException, proxy_server.openai_exception_handler) + if caller is not None: + usage_cache: Final = InternalUsageCache(cache) + configured_limiter: Final = ( + _PROXY_MaxParallelRequestsHandler_v3(usage_cache) if isinstance(limiter, str) else limiter + ) + monkeypatch.setattr(proxy_server, "proxy_logging_obj", _ProxyLogging(usage_cache, configured_limiter)) + app.dependency_overrides[endpoint.user_api_key_auth] = lambda: caller + return app + + +async def _post( + app: FastAPI, + body: Mapping[str, JsonValue], + *, + current_deployment_id: str = "opus", + candidate_deployment_id: str = "sonnet", +) -> httpx.Response: + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + return await client.post( + "/cost/predict-cache", + json={ + "current_deployment_id": current_deployment_id, + "candidate_deployment_id": candidate_deployment_id, + "request": body, + }, + ) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("warm_deployment", "warm_model", "expected_delta", "expected_penalty"), + [("sonnet", "claude-sonnet-5", -0.03325, 0.0), ("opus", "claude-opus-5", 0.007, 0.0115)], +) +async def test_switch_delta_accounts_for_each_deployment_cache( + monkeypatch: pytest.MonkeyPatch, + warm_deployment: str, + warm_model: str, + expected_delta: float, + expected_penalty: float, +) -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body, deployment_id=warm_deployment, model=warm_model) + app: Final = _app(monkeypatch, cache, caller=UserAPIKeyAuth(api_key=_CALLER)) + response: Final = await _post(app, body) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.switch_delta == pytest.approx(expected_delta) + assert result.cache_rebuild_penalty == pytest.approx(expected_penalty) + assert result.cache_guarantee is False + assert result.pricing_basis == "input_before_discounts_and_margins" + if warm_deployment == "sonnet": + assert result.switch.cache_state == "warm" + assert result.stay.cache_state == "unknown" + else: + assert result.stay.cache_state == "warm" + assert result.switch.cache_state == "unknown" + + +@pytest.mark.asyncio +async def test_missing_caller_identity_cannot_reuse_observations(monkeypatch: pytest.MonkeyPatch) -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body) + response: Final = await _post( + _app(monkeypatch, cache, caller=UserAPIKeyAuth(api_key=None), counts=_unexpected_count), body + ) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.reason == result.switch.reason == "caller_identity_unavailable" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None + + +@pytest.mark.asyncio +async def test_unauthenticated_request_is_rejected(monkeypatch: pytest.MonkeyPatch) -> None: + import litellm.proxy.proxy_server as proxy_server + + monkeypatch.setattr(proxy_server, "master_key", "cache-prediction-test-master-key") + response: Final = await _post(_app(monkeypatch, DualCache()), _body()) + assert response.status_code == 401, response.text + + +@pytest.mark.asyncio +@pytest.mark.parametrize("arm", ["current", "candidate"]) +@pytest.mark.parametrize("caller_team", [None, "own-team"]) +@pytest.mark.parametrize("restricted", [False, True]) +async def test_foreign_and_missing_deployments_have_identical_authenticated_responses( + monkeypatch: pytest.MonkeyPatch, arm: str, caller_team: str | None, restricted: bool +) -> None: + allowed: Final = ("sonnet",) if arm == "current" else ("opus",) + app: Final = _app( + monkeypatch, + DualCache(), + caller=UserAPIKeyAuth(api_key=_CALLER, team_id=caller_team, models=list(allowed) if restricted else []), + current_team="foreign-team" if arm == "current" else None, + candidate_team="foreign-team" if arm == "candidate" else None, + counts=_unexpected_count, + ) + foreign: Final = await _post(app, _body()) + missing: Final = await _post( + app, + _body(), + current_deployment_id="missing-deployment" if arm == "current" else "opus", + candidate_deployment_id="missing-deployment" if arm == "candidate" else "sonnet", + ) + + assert foreign.status_code == missing.status_code == 404 + assert foreign.json() == missing.json() == {"detail": "Deployment not found"} + + +@pytest.mark.asyncio +@pytest.mark.parametrize("deployment_team", [None, "own-team"]) +async def test_visible_public_and_own_team_deployments_remain_available( + monkeypatch: pytest.MonkeyPatch, deployment_team: str | None +) -> None: + app: Final = _app( + monkeypatch, + DualCache(), + caller=UserAPIKeyAuth(api_key=_CALLER, team_id="own-team"), + current_team=deployment_team, + candidate_team=deployment_team, + ) + response: Final = await _post(app, _body()) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.estimate is not None and result.switch.estimate is not None + + +@pytest.mark.asyncio +@pytest.mark.parametrize("arm", ["current", "candidate"]) +async def test_visible_deployment_outside_key_model_permissions_is_forbidden( + monkeypatch: pytest.MonkeyPatch, arm: str +) -> None: + allowed: Final = "sonnet" if arm == "current" else "opus" + denied: Final = "opus" if arm == "current" else "sonnet" + app: Final = _app(monkeypatch, DualCache(), caller=UserAPIKeyAuth(api_key=_CALLER, models=[allowed])) + response: Final = await _post(app, _body()) + assert response.status_code == 403, response.text + assert denied in response.text + + +@pytest.mark.asyncio +async def test_other_callers_warm_cache_is_not_prediction_evidence(monkeypatch: pytest.MonkeyPatch) -> None: + cache: Final = DualCache() + body: Final = _body() + await _observe(cache, body, caller="other-caller") + response: Final = await _post(_app(monkeypatch, cache, caller=UserAPIKeyAuth(api_key=_CALLER)), body) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.switch.cache_state == "unknown" + assert result.switch.reason == "no_compatible_observation" + assert result.switch.evidence is None + assert result.switch.estimate is not None + assert result.switch.estimate.tokens.cache_read_input_tokens == 0 + + +@pytest.mark.asyncio +async def test_count_failure_nulls_switch_comparison(monkeypatch: pytest.MonkeyPatch) -> None: + app: Final = _app( + monkeypatch, DualCache(), caller=UserAPIKeyAuth(api_key=_CALLER), counts=Counts(total=None) + ) + response: Final = await _post(app, _body()) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.reason == result.switch.reason == "token_count_unavailable" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None + + +@pytest.mark.asyncio +@pytest.mark.parametrize("limiter", [None, CustomLogger()]) +async def test_missing_or_unsupported_limiter_returns_unknown_before_counting( + monkeypatch: pytest.MonkeyPatch, limiter: CustomLogger | None +) -> None: + app: Final = _app( + monkeypatch, DualCache(), caller=UserAPIKeyAuth(api_key=_CALLER), counts=_unexpected_count, limiter=limiter + ) + response: Final = await _post(app, _body()) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.reason == result.switch.reason == "limiter_unavailable" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None + + +@pytest.mark.asyncio +async def test_occupied_parallel_capacity_rejects_before_provider_count(monkeypatch: pytest.MonkeyPatch) -> None: + cache: Final = DualCache() + limiter: Final = _PROXY_MaxParallelRequestsHandler_v3(InternalUsageCache(cache)) + caller: Final = UserAPIKeyAuth(api_key=_CALLER, max_parallel_requests=1) + app: Final = _app(monkeypatch, cache, caller=caller, counts=_unexpected_count, limiter=limiter) + async with limiter.request_capacity(caller, "opus"): + response: Final = await _post(app, _body()) + + assert response.status_code == 429, response.text + assert "max_parallel_requests" in response.text + recovered: Final = await _post(_app(monkeypatch, cache, caller=caller, limiter=limiter), _body()) + assert recovered.status_code == 200, recovered.text + + +@pytest.mark.asyncio +async def test_each_count_consumes_the_deployment_group_rpm_limit(monkeypatch: pytest.MonkeyPatch) -> None: + calls: Final = asyncio.Queue[str]() + + async def count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + calls.put_nowait(model) + return await Counts()(model, api_key, body) + + caller: Final = UserAPIKeyAuth(api_key=_CALLER, metadata={"model_rpm_limit": {"sonnet": 1}}) + app: Final = _app(monkeypatch, DualCache(), caller=caller, counts=count) + response: Final = await _post(app, _body()) + + assert response.status_code == 429, response.text + assert calls.qsize() == 3 + assert tuple(calls.get_nowait() for _ in range(3)) == ( + "claude-opus-5", "claude-opus-5", "claude-sonnet-5" + ) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) +async def test_each_count_preserves_auth_cached_request_tag_limits( + monkeypatch: pytest.MonkeyPatch, metadata_key: str +) -> None: + calls: Final = asyncio.Queue[str]() + caller: Final = UserAPIKeyAuth(api_key=_CALLER, metadata={"tag_rpm_limit": {"cache-cost": 1}}) + + async def count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + calls.put_nowait(model) + return await Counts()(model, api_key, body) + + async def authenticated_request(request: Request) -> UserAPIKeyAuth: + data: Final = await _read_request_body(request) + _safe_set_request_parsed_body(request, {**data, metadata_key: {"tags": ["cache-cost"]}}) + return caller + + app: Final = _app(monkeypatch, DualCache(), caller=caller, counts=count) + app.dependency_overrides[endpoint.user_api_key_auth] = authenticated_request + response: Final = await _post(app, _body()) + + assert response.status_code == 429, response.text + assert "tag_per_key" in response.text + assert calls.qsize() == 1 + assert calls.get_nowait() == "claude-opus-5" + + +@pytest.mark.asyncio +async def test_provider_counter_failure_releases_parallel_capacity(monkeypatch: pytest.MonkeyPatch) -> None: + cache: Final = DualCache() + limiter: Final = _PROXY_MaxParallelRequestsHandler_v3(InternalUsageCache(cache)) + caller: Final = UserAPIKeyAuth(api_key=_CALLER, max_parallel_requests=1) + + async def fail_count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + raise RuntimeError("provider counter failed") + + app: Final = _app(monkeypatch, cache, caller=caller, counts=fail_count, limiter=limiter) + with pytest.raises(RuntimeError, match="provider counter failed"): + await _post(app, _body()) + recovered: Final = await _post(_app(monkeypatch, cache, caller=caller, limiter=limiter), _body()) + assert recovered.status_code == 200, recovered.text + assert recovered.json()["switch"]["estimate"]["input_cost"] == pytest.approx(0.0145) + + +@pytest.mark.asyncio +async def test_cancelled_provider_counter_releases_parallel_capacity(monkeypatch: pytest.MonkeyPatch) -> None: + cache: Final = DualCache() + limiter: Final = _PROXY_MaxParallelRequestsHandler_v3(InternalUsageCache(cache)) + caller: Final = UserAPIKeyAuth(api_key=_CALLER, max_parallel_requests=1) + started: Final = asyncio.Event() + release: Final = asyncio.Event() + + async def wait_count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + started.set() + await release.wait() + return await Counts()(model, api_key, body) + + app: Final = _app(monkeypatch, cache, caller=caller, counts=wait_count, limiter=limiter) + pending: Final = asyncio.create_task(_post(app, _body())) + try: + await asyncio.wait_for(started.wait(), timeout=5) + pending.cancel() + with pytest.raises(asyncio.CancelledError): + await pending + release.set() + recovered: Final = await asyncio.wait_for(_post(app, _body()), timeout=5) + assert recovered.status_code == 200, recovered.text + assert recovered.json()["switch"]["estimate"]["input_cost"] == pytest.approx(0.0145) + finally: + pending.cancel() + release.set() + await asyncio.gather(pending, return_exceptions=True) + + +async def _unexpected_count(model: str, api_key: str, body: Mapping[str, JsonValue]) -> int | None: + pytest.fail("Unsupported prediction must return before contacting the token counter") + + +class RequestMutator(CustomLogger): + async def async_pre_call_hook( + self, user_api_key_dict: UserAPIKeyAuth, cache: DualCache, data: dict[str, object], call_type: str + ) -> dict[str, object]: + return {**data, "system": "Injected policy"} + + +@pytest.fixture +def request_mutator() -> Iterator[RequestMutator]: + callback: Final = RequestMutator() + litellm.logging_callback_manager.add_litellm_callback(callback) + try: + yield callback + finally: + litellm.logging_callback_manager.remove_callback_from_all_lists(callback) + + +@pytest.mark.asyncio +async def test_request_transform_callback_returns_unknown_before_token_counting( + monkeypatch: pytest.MonkeyPatch, request_mutator: RequestMutator +) -> None: + app: Final = _app( + monkeypatch, DualCache(), caller=UserAPIKeyAuth(api_key=_CALLER), counts=_unexpected_count + ) + response: Final = await _post(app, _body()) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.cache_state == result.switch.cache_state == "unknown" + assert result.stay.reason == result.switch.reason == "unsupported_request_transform" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None + + +@pytest.mark.asyncio +async def test_key_config_returns_unknown_before_token_counting(monkeypatch: pytest.MonkeyPatch) -> None: + app: Final = _app( + monkeypatch, + DualCache(), + caller=UserAPIKeyAuth(api_key=_CALLER, config={"model_list": []}), + counts=_unexpected_count, + ) + response: Final = await _post(app, _body()) + + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.cache_state == result.switch.cache_state == "unknown" + assert result.stay.reason == result.switch.reason == "unsupported_request_transform" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None + + +@pytest.mark.parametrize("headers", [ + {"anthropic-version": "2099-01-01"}, + {"anthropic-beta": "future-feature"}, +]) +@pytest.mark.asyncio +async def test_unsupported_provider_headers_cannot_reuse_default_version_evidence( + monkeypatch: pytest.MonkeyPatch, headers: dict[str, str] +) -> None: + cache: Final = DualCache() + await _observe(cache, _body(), deployment_id="sonnet") + app: Final = _app( + monkeypatch, cache, caller=UserAPIKeyAuth(api_key=_CALLER), counts=_unexpected_count + ) + async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client: + response: Final = await client.post( + "/cost/predict-cache", + headers=headers, + json={"current_deployment_id": "opus", "candidate_deployment_id": "sonnet", "request": _body()}, + ) + assert response.status_code == 200, response.text + result: Final = CachePredictionResponse.model_validate(response.json()) + assert result.stay.cache_state == result.switch.cache_state == "unknown" + assert result.stay.reason == result.switch.reason == "unsupported_provider_headers" + assert result.stay.estimate is None and result.switch.estimate is None + assert result.switch_delta is None and result.cache_rebuild_penalty is None diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 9e844b992b2..cac1005cc2f 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -3387,6 +3387,35 @@ export interface paths { patch?: never; trace?: never; }; + "/cost/predict-cache": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + get?: never; + put?: never; + /** + * Predict Cache Cost + * @description Compare the next native Anthropic request on two configured deployment IDs. + * + * Estimates use provider token counting and recent successful cache telemetry for this key. + * Unknown cache state uses the cold scenario when prices/counts are available. Cache observations + * do not guarantee retention. v0 supports one message-content breakpoint, text and client tools; + * system/tool-only breakpoints, thinking, images, nondefault Anthropic versions, beta headers and + * request transforms are unknown. + * Each provider count consumes one RPM unit and holds concurrency capacity; a comparison uses + * up to four counts. The legacy rate limiter returns unknown without contacting the provider. + * This endpoint does not generate tokens, prewarm caches, choose a model or alter routing. + */ + post: operations["predict_cache_cost_cost_predict_cache_post"]; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; "/credentials": { parameters: { query?: never; @@ -24618,6 +24647,31 @@ export interface components { /** Failed Requests */ failed_requests: number; }; + /** CacheCostScenario */ + CacheCostScenario: { + /** Input Cost */ + input_cost: number; + tokens: components["schemas"]["CacheTokenBuckets"]; + }; + /** CacheEvidence */ + CacheEvidence: { + /** + * Confidence + * @default observed + * @constant + */ + confidence: "observed"; + /** Expires At */ + expires_at: number; + /** Observed At */ + observed_at: number; + /** + * Source + * @default provider_usage + * @constant + */ + source: "provider_usage"; + }; /** CachePingResponse */ CachePingResponse: { /** Cache Type */ @@ -24635,6 +24689,59 @@ export interface components { /** Status */ status: string; }; + /** CachePredictionArm */ + CachePredictionArm: { + /** + * Cache State + * @default unknown + * @enum {string} + */ + cache_state: "warm" | "partial" | "stale" | "unknown" | "disabled"; + cold?: components["schemas"]["CacheCostScenario"] | null; + /** Deployment Id */ + deployment_id: string; + estimate?: components["schemas"]["CacheCostScenario"] | null; + evidence?: components["schemas"]["CacheEvidence"] | null; + /** Model */ + model?: string | null; + /** Reason */ + reason?: string | null; + /** Token Count Source */ + token_count_source?: "anthropic_count_tokens" | null; + warm?: components["schemas"]["CacheCostScenario"] | null; + }; + /** CachePredictionRequest */ + CachePredictionRequest: { + /** Candidate Deployment Id */ + candidate_deployment_id: string; + /** Current Deployment Id */ + current_deployment_id: string; + /** Request */ + request: { + [key: string]: components["schemas"]["JsonValue"]; + }; + }; + /** CachePredictionResponse */ + CachePredictionResponse: { + /** + * Cache Guarantee + * @default false + * @constant + */ + cache_guarantee: false; + /** Cache Rebuild Penalty */ + cache_rebuild_penalty: number | null; + /** + * Pricing Basis + * @default input_before_discounts_and_margins + * @constant + */ + pricing_basis: "input_before_discounts_and_margins"; + stay: components["schemas"]["CachePredictionArm"]; + switch: components["schemas"]["CachePredictionArm"]; + /** Switch Delta */ + switch_delta: number | null; + }; /** CacheSettingsField */ CacheSettingsField: { /** Field Default */ @@ -24716,6 +24823,29 @@ export interface components { */ status: string; }; + /** CacheTokenBuckets */ + CacheTokenBuckets: { + /** + * Cache Creation 1H Input Tokens + * @default 0 + */ + cache_creation_1h_input_tokens: number; + /** + * Cache Creation 5M Input Tokens + * @default 0 + */ + cache_creation_5m_input_tokens: number; + /** + * Cache Read Input Tokens + * @default 0 + */ + cache_read_input_tokens: number; + /** + * Uncached Input Tokens + * @default 0 + */ + uncached_input_tokens: number; + }; /** * CallTypes * @enum {string} @@ -28332,6 +28462,7 @@ export interface components { /** Updated By */ updated_by?: string | null; }; + JsonValue: unknown; /** KeyHealthResponse */ KeyHealthResponse: { /** @@ -45167,6 +45298,39 @@ export interface operations { }; }; }; + predict_cache_cost_cost_predict_cache_post: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody: { + content: { + "application/json": components["schemas"]["CachePredictionRequest"]; + }; + }; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["CachePredictionResponse"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; get_credentials_credentials_get: { parameters: { query?: never; From 555e321cf170ad2d15a2932fccdea104da1c39b6 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 01:05:51 +0000 Subject: [PATCH 006/119] fix(router): record flat retry attempts and cap retries from attempted_retries Router.log_retry used to copy the failed attempt's kwargs and metadata into metadata.previous_models. Nothing downstream read those copies, but they carried client credentials into spend logs and grew the payload on every retry. Each attempt now leaves a flat record (model group, deployment id, exception type and string, attempt number), which drops RETRY_BREADCRUMB_EXCLUDED_KWARGS and the per-retry credential masking. num_retries_per_request was enforced from len(previous_models), which only looked at the metadata bucket and never exceeded four records. The sync and async client wrappers and the Rust lifecycle guard now read attempted_retries from whichever metadata bucket the call carries. Resolves LIT-7505 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/__init__.py | 2 +- litellm/litellm_core_utils/core_helpers.py | 13 +++ litellm/router.py | 45 +++----- litellm/rust_bridge/lifecycle.py | 16 +-- litellm/types/router.py | 8 ++ litellm/utils.py | 18 +-- .../test_router_helper_utils.py | 29 +++-- .../rust_bridge/test_lifecycle.py | 30 +++++ tests/test_litellm/test_router.py | 106 +++++++++++------- tests/test_litellm/test_utils.py | 47 ++++++++ tests/test_litellm_rust/ocr/test_lifecycle.py | 2 +- 11 files changed, 209 insertions(+), 107 deletions(-) create mode 100644 tests/test_litellm/rust_bridge/test_lifecycle.py diff --git a/litellm/__init__.py b/litellm/__init__.py index ccfbf80369f..261457d6889 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -538,7 +538,7 @@ context_window_fallbacks: Optional[List] = None content_policy_fallbacks: Optional[List] = None allowed_fails: int = 3 allow_dynamic_callback_disabling: bool = True -num_retries_per_request: Optional[int] = None # for the request overall (incl. fallbacks + model retries) +num_retries_per_request: Optional[int] = None # cap on Router retries of one model group; resets per fallback hop ####### SECRET MANAGERS ##################### secret_manager_client: Optional[Any] = ( None # list of instantiated key management clients - e.g. azure kv, infisical, etc. diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py index aa7d6ca1699..ecd9cdac88b 100644 --- a/litellm/litellm_core_utils/core_helpers.py +++ b/litellm/litellm_core_utils/core_helpers.py @@ -303,6 +303,19 @@ def get_metadata_variable_name_from_kwargs( return "litellm_metadata" if "litellm_metadata" in kwargs else "metadata" +def max_retries_per_request_hit(kwargs: Mapping[str, object], num_retries_per_request: int | None) -> bool: + """ + Whether the Router retry about to run (``attempted_retries`` >= 1 in the metadata bucket) is past the cap + """ + if num_retries_per_request is None: + return False + metadata: Final = kwargs.get(get_metadata_variable_name_from_kwargs(kwargs)) + if not isinstance(metadata, Mapping): + return False + attempted_retries: Final = metadata.get("attempted_retries") + return type(attempted_retries) is int and 0 < attempted_retries and num_retries_per_request <= attempted_retries + + def get_or_create_metadata_bucket( request_data: dict, ) -> tuple[Literal["metadata", "litellm_metadata"], dict]: diff --git a/litellm/router.py b/litellm/router.py index 8865543badd..a46ffa83b05 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -96,7 +96,6 @@ from litellm.litellm_core_utils.request_timeout_resolver import ( from litellm.litellm_core_utils.secret_redaction import redact_string from litellm.litellm_core_utils.sensitive_data_masker import ( SensitiveDataMasker, - mask_credentials_in_payload, mask_sensitive_structure, ) from litellm.litellm_core_utils.token_counter import offload_token_count @@ -242,6 +241,7 @@ from litellm.types.router import ( ModelGroupInfo, OptionalPreCallChecks, PreRoutingStrategy, + RetryAttemptRecord, RetryPolicy, RouterCacheEnum, RouterErrors, @@ -623,20 +623,6 @@ def _replay_live_router_model_cost() -> None: set_live_deployment_replay(_replay_live_router_model_cost) -# Kwargs that carry no signal about the failed attempt, so log_retry drops them from a -# breadcrumb entirely: the request payload, the proxy's snapshot of the inbound request (its body -# aliases the live request metadata, earlier breadcrumbs included, so copying it would nest every -# breadcrumb inside the next one), and the router-internal walk state. Credentials are handled -# separately by mask_credentials_in_payload, which scrubs credential-named values from whatever -# kwargs remain rather than trying to enumerate every credential-bearing key here. -RETRY_BREADCRUMB_EXCLUDED_KWARGS: Final = frozenset( - ( - "messages", - "original_function", - "attempted_targets", - "proxy_server_request", - ) -) RETRY_BREADCRUMB_LIMIT: Final = 4 @@ -8374,31 +8360,28 @@ class Router: def log_retry(self, kwargs: dict, e: Exception) -> dict: """ - When a retry or fallback happens, log the details of the just failed model call - similar to Sentry breadcrumbing + When a retry or fallback happens, record which model group, deployment and attempt just failed and why """ _metadata_var: Final = "litellm_metadata" if "litellm_metadata" in kwargs else "metadata" request_metadata: Final[Mapping[str, object]] = kwargs[_metadata_var] - attempt_kwargs: Final = MappingProxyType( - {k: v for k, v in kwargs.items() if k != _metadata_var and k not in RETRY_BREADCRUMB_EXCLUDED_KWARGS} - ) - attempt_metadata: Final = MappingProxyType( - {k: v for k, v in request_metadata.items() if k != "previous_models"} - ) - previous_model: Final = MappingProxyType( - { - "exception_type": type(e).__name__, - "exception_string": str(e), - **attempt_kwargs, - _metadata_var: attempt_metadata, - } - ) + model_group: Final = kwargs.get("model") + model_info: Final = request_metadata.get("model_info") + deployment_id: Final = model_info.get("id") if isinstance(model_info, Mapping) else None + attempted_retries: Final = request_metadata.get("attempted_retries") + attempt_record: Final[RetryAttemptRecord] = { + "model_group": model_group if isinstance(model_group, str) else None, + "deployment_id": deployment_id if isinstance(deployment_id, str) else None, + "exception_type": type(e).__name__, + "exception_string": str(e), + "attempted_retries": attempted_retries if type(attempted_retries) is int else None, + } earlier_breadcrumbs: Final = request_metadata.get("previous_models") kept_breadcrumbs: Final[tuple[object, ...]] = ( tuple(earlier_breadcrumbs)[-(RETRY_BREADCRUMB_LIMIT - 1) :] if isinstance(earlier_breadcrumbs, (list, tuple)) else () ) - breadcrumbs: Final = (*kept_breadcrumbs, mask_credentials_in_payload(previous_model)) + breadcrumbs: Final = (*kept_breadcrumbs, attempt_record) kwargs[_metadata_var]["previous_models"] = breadcrumbs # rebind-ok: the logging object already holds this dict return kwargs diff --git a/litellm/rust_bridge/lifecycle.py b/litellm/rust_bridge/lifecycle.py index f5e0c1b0fc6..f1cc912129d 100644 --- a/litellm/rust_bridge/lifecycle.py +++ b/litellm/rust_bridge/lifecycle.py @@ -99,23 +99,13 @@ def setup( def check_limits(kwargs: Mapping[str, object]) -> None: import litellm + from litellm.litellm_core_utils.core_helpers import max_retries_per_request_hit current_cost: Final = litellm._current_cost # pyright: ignore[reportPrivateUsage] # shared SDK budget counter has no public accessor if litellm.max_budget and current_cost > litellm.max_budget: raise litellm.BudgetExceededError(current_cost=current_cost, max_budget=litellm.max_budget) - metadata: Final = kwargs.get("metadata") - if isinstance(metadata, Mapping): - typed_metadata: Final = cast( # cast-ok: runtime Mapping check establishes read-only metadata - Mapping[str, object], metadata - ) - previous: Final = typed_metadata.get("previous_models") - if ( - isinstance(previous, list) - and litellm.num_retries_per_request is not None - and len(cast(list[object], previous)) # cast-ok: runtime list check establishes the retry history - >= litellm.num_retries_per_request - ): - raise RuntimeError("Max retries per request hit!") + if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request): + raise RuntimeError("Max retries per request hit!") def finalize( diff --git a/litellm/types/router.py b/litellm/types/router.py index fc09c40fe08..fecc0e00f99 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -883,6 +883,14 @@ class RouterModelGroupAliasItem(TypedDict): hidden: bool # if 'True', don't return on `.get_model_list` +class RetryAttemptRecord(TypedDict): + model_group: ReadOnly[str | None] + deployment_id: ReadOnly[str | None] + exception_type: ReadOnly[str] + exception_string: ReadOnly[str] + attempted_retries: ReadOnly[int | None] + + VALID_LITELLM_ENVIRONMENTS = [ "development", "staging", diff --git a/litellm/utils.py b/litellm/utils.py index 394ab4b4094..98881e68986 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -81,7 +81,7 @@ from litellm.constants import ( PROVIDERS_THAT_AUTHENTICATE_ON_PROVIDER_INFO, TOOL_CHOICE_OBJECT_TOKEN_COUNT, ) -from litellm.litellm_core_utils.core_helpers import normalize_drop_params +from litellm.litellm_core_utils.core_helpers import max_retries_per_request_hit, normalize_drop_params from litellm.litellm_core_utils.fallback_generalizations import ( match_capability_generalizations, ) @@ -1509,12 +1509,8 @@ def client(original_function): call_type = original_function.__name__ if _is_async_request(kwargs): # [OPTIONAL] CHECK MAX RETRIES / REQUEST - if litellm.num_retries_per_request is not None: - # check if previous_models passed in as ['litellm_params']['metadata]['previous_models'] - previous_models = (kwargs.get("metadata") or {}).get("previous_models", None) - if previous_models is not None: - if litellm.num_retries_per_request <= len(previous_models): - raise Exception("Max retries per request hit!") + if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request): + raise Exception("Max retries per request hit!") # MODEL CALL result = original_function(*args, **kwargs) @@ -1573,12 +1569,8 @@ def client(original_function): ) # [OPTIONAL] CHECK MAX RETRIES / REQUEST - if litellm.num_retries_per_request is not None: - # check if previous_models passed in as ['litellm_params']['metadata]['previous_models'] - previous_models = (kwargs.get("metadata") or {}).get("previous_models", None) - if previous_models is not None: - if litellm.num_retries_per_request <= len(previous_models): - raise Exception("Max retries per request hit!") + if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request): + raise Exception("Max retries per request hit!") # [OPTIONAL] CHECK CACHE print_verbose( diff --git a/tests/router_unit_tests/test_router_helper_utils.py b/tests/router_unit_tests/test_router_helper_utils.py index 7dbac243d55..5b06c5fdb01 100644 --- a/tests/router_unit_tests/test_router_helper_utils.py +++ b/tests/router_unit_tests/test_router_helper_utils.py @@ -1,3 +1,4 @@ +import json import os import traceback from dotenv import load_dotenv @@ -628,17 +629,29 @@ def test_deployment_callback_respects_cooldown_time(model_list): assert mock_set.call_args.kwargs["time_to_cooldown"] == 0 -def test_log_retry(model_list): - """Test if the '_log_retry' function is working correctly""" - import time - +@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) +def test_log_retry(model_list, metadata_key): + """log_retry appends one flat record per failed attempt and copies neither the request kwargs nor + the request metadata into it""" router = Router(model_list=model_list) new_kwargs = router.log_retry( - kwargs={"metadata": {}}, - e=Exception(), + kwargs={ + "model": "gpt-3.5-turbo", + "api_key": "sk-must-not-be-recorded", + "messages": [{"role": "user", "content": "hi"}], + metadata_key: {"model_info": {"id": "deployment-1"}, "attempted_retries": 2, "user_api_key": "sk-proxy"}, + }, + e=litellm.RateLimitError(message="slow down", llm_provider="openai", model="gpt-3.5-turbo"), ) - assert "metadata" in new_kwargs - assert "previous_models" in new_kwargs["metadata"] + assert json.loads(json.dumps(new_kwargs[metadata_key]["previous_models"])) == [ + { + "model_group": "gpt-3.5-turbo", + "deployment_id": "deployment-1", + "exception_type": "RateLimitError", + "exception_string": "litellm.RateLimitError: slow down", + "attempted_retries": 2, + } + ] def test_update_usage(model_list): diff --git a/tests/test_litellm/rust_bridge/test_lifecycle.py b/tests/test_litellm/rust_bridge/test_lifecycle.py new file mode 100644 index 00000000000..1f0b5591c2b --- /dev/null +++ b/tests/test_litellm/rust_bridge/test_lifecycle.py @@ -0,0 +1,30 @@ +from typing import Final + +import pytest + +import litellm +from litellm.rust_bridge.lifecycle import check_limits + + +@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) +@pytest.mark.parametrize( + "cap, attempted_retries, refused", + [(5, 5, True), (5, 4, False), (0, 0, False), (0, 1, True)], + ids=[ + "cap-above-four-reached", + "cap-above-four-not-reached", + "first-attempt-passes-cap-of-zero", + "cap-of-zero-refuses-first-retry", + ], +) +def test_check_limits_reads_attempted_retries( + monkeypatch: pytest.MonkeyPatch, metadata_key: str, cap: int, attempted_retries: int, refused: bool +) -> None: + monkeypatch.setattr(litellm, "num_retries_per_request", cap) + monkeypatch.setattr(litellm, "max_budget", None) + kwargs: Final = {"model": "mistral/mistral-ocr-latest", metadata_key: {"attempted_retries": attempted_retries}} + if refused: + with pytest.raises(RuntimeError, match="Max retries per request hit!"): + check_limits(kwargs) + else: + check_limits(kwargs) diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index f5e9b2091a0..6150ce287ed 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -10607,6 +10607,7 @@ def _cyclic_fallback_router(num_retries=0): "api_key": "sk-fake", "mock_response": "litellm.InternalServerError", }, + "model_info": {"id": f"{group}-deployment"}, } for group in groups ], @@ -10656,28 +10657,37 @@ async def test_cyclic_fallback_graph_does_not_amplify_one_request(): assert sum(len(message) for message in capture.messages) < 5_000 +_FLAT_ATTEMPT_RECORD_KEYS = frozenset( + {"model_group", "deployment_id", "exception_type", "exception_string", "attempted_retries"} +) +_BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doNotShip" + + @pytest.mark.asyncio -async def test_retry_breadcrumbs_do_not_carry_the_walk_state(): - """log_retry copies every kwarg into previous_models, which reaches spend logs and - logging callbacks. The set of already-attempted groups is router-internal walk state - with no diagnostic value there, and it is the one entry that is not a plain scalar. - A retry has to be configured for the walk state to reach log_retry at all.""" +async def test_retry_records_are_flat_and_name_the_failed_group_on_fallback_hops(): + """Each failed attempt leaves a flat record in previous_models, which reaches spend logs and + logging callbacks. Nothing downstream reads the failed attempt's kwargs or metadata, and copying + them is what carried client credentials and multiplied the payload on every retry. A fallback hop + calls log_retry too, so the record has to name the group that failed, not the one taken next.""" router = _cyclic_fallback_router(num_retries=1) capture = _LogCapture(logging.ERROR) recorder = _FallbackAttemptRecorder() await _drive_cyclic_fallback(router, capture, recorder) - breadcrumbs = [breadcrumb for hop in recorder.breadcrumbs_per_target for breadcrumb in hop] - assert breadcrumbs, "no retry breadcrumbs were recorded" - assert any( - "fallback_depth" in breadcrumb for breadcrumb in breadcrumbs - ), "no breadcrumb carried router walk state, so this test cannot see the leak" - for breadcrumb in breadcrumbs: - assert "attempted_targets" not in breadcrumb - - -_BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doNotShip" + records = [record for hop in recorder.breadcrumbs_per_target for record in hop] + assert records, "no retry records were recorded" + for record in records: + assert set(record) == _FLAT_ATTEMPT_RECORD_KEYS + assert record["exception_type"] == "InternalServerError" + assert record["deployment_id"] == f"{record['model_group']}-deployment" + group_failed_before_hop = {"group-b": "group-a", "group-c": "group-b", "group-d": "group-c"} + for failed_target, hop_records in zip(recorder.failed_targets, recorder.breadcrumbs_per_target): + groups = [record["model_group"] for record in hop_records] + first_own_attempt = groups.index(failed_target) + assert groups[first_own_attempt - 1] == group_failed_before_hop[failed_target] + assert set(groups[first_own_attempt:]) == {failed_target} + assert [record["attempted_retries"] for record in hop_records[first_own_attempt:]][:2] == [0, 1] @pytest.mark.parametrize( @@ -10703,22 +10713,20 @@ _BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doN ], ) @pytest.mark.asyncio -async def test_retry_breadcrumbs_never_carry_a_forwarded_credential(container_key, request_kwargs): - """log_retry copies kwargs into previous_models, which reaches spend logs and logging callbacks. - Any of these kwargs can carry a client's forwarded Authorization token or a provider key, and a - breadcrumb has no diagnostic use for the raw secret. A denylist of key names is always one new - credential kwarg behind, so log_retry scrubs credential-named values by pattern instead: the - container still reaches the breadcrumb, but the raw secret never does, whatever key holds it.""" +async def test_retry_records_never_carry_a_forwarded_credential(container_key, request_kwargs): + """previous_models reaches spend logs and logging callbacks. Any request kwarg can carry a client's + forwarded Authorization token or a provider key, so the record must not carry request kwargs at + all: neither the credential-bearing container nor the raw secret, whatever key holds it.""" router = _cyclic_fallback_router(num_retries=1) capture = _LogCapture(logging.ERROR) metadata = {} await _drive_cyclic_fallback(router, capture, metadata=metadata, **request_kwargs) - breadcrumbs = metadata["previous_models"] - assert breadcrumbs, "no retry breadcrumbs were recorded" - dumped = json.dumps(breadcrumbs, default=str) - assert container_key in dumped, "the credential-bearing kwarg never reached the breadcrumb, so this test cannot see the leak" + records = metadata["previous_models"] + assert records, "no retry records were recorded" + dumped = json.dumps(records) + assert container_key not in dumped assert _BREADCRUMB_CREDENTIAL_CANARY not in dumped @@ -10743,7 +10751,7 @@ async def _fail_one_proxy_shaped_request(router, request_marker): shallow copy of the request, so body["metadata"] is the very same dict the router later stamps previous_models onto.""" metadata = {"request_marker": request_marker} - with pytest.raises(litellm.InternalServerError): + with pytest.raises((litellm.InternalServerError, litellm.APIConnectionError)): await router.acompletion( model="broken-group", messages=[{"role": "user", "content": "hi"}], @@ -10769,34 +10777,52 @@ def _nested_breadcrumb_lists(node): @pytest.mark.asyncio -async def test_retry_breadcrumbs_stay_per_request_and_flat_across_failing_requests(): - """Every failed attempt appends a breadcrumb to metadata["previous_models"], and the proxy's +async def test_retry_records_stay_per_request_and_flat_across_failing_requests(): + """Every failed attempt appends a record to metadata["previous_models"], and the proxy's request snapshot aliases that same metadata dict. Kept on the Router and copied wholesale, - each breadcrumb embedded every earlier one from every earlier request, so the breadcrumb + each breadcrumb once embedded every earlier one from every earlier request, so the breadcrumb tree, and with it the debug repr of the kwargs, roughly doubled on each failed attempt until a single-worker proxy spent minutes in the redaction regex and stopped answering.""" router = _always_failing_router(num_retries=2) - breadcrumbs_per_request = [ + records_per_request = [ await _fail_one_proxy_shaped_request(router, f"request-{request_number}") for request_number in range(1, 7) ] - for request_number, breadcrumbs in enumerate(breadcrumbs_per_request, start=1): - assert len(breadcrumbs) == 3, "one initial attempt plus two retries failed, each leaving one breadcrumb" - assert {breadcrumb["metadata"]["request_marker"] for breadcrumb in breadcrumbs} == {f"request-{request_number}"} - for breadcrumb in breadcrumbs: - assert _nested_breadcrumb_lists(breadcrumb) == [] - assert len({len(repr(breadcrumbs)) for breadcrumbs in breadcrumbs_per_request}) == 1 + for records in records_per_request: + assert [record["attempted_retries"] for record in records] == [0, 1, 2] + for record in records: + assert set(record) == _FLAT_ATTEMPT_RECORD_KEYS + assert _nested_breadcrumb_lists(record) == [] + assert len({len(repr(records)) for records in records_per_request}) == 1 @pytest.mark.asyncio -async def test_retry_breadcrumbs_keep_only_the_last_four_attempts(): +async def test_retry_records_keep_only_the_last_four_attempts(): router = _always_failing_router(num_retries=6) - breadcrumbs = await _fail_one_proxy_shaped_request(router, "request-1") + records = await _fail_one_proxy_shaped_request(router, "request-1") - assert len(breadcrumbs) == 4 - assert [breadcrumb["metadata"]["attempted_retries"] for breadcrumb in breadcrumbs] == [3, 4, 5, 6] + assert [record["attempted_retries"] for record in records] == [3, 4, 5, 6] + + +@pytest.mark.asyncio +async def test_num_retries_per_request_stops_retries_at_caps_above_four(monkeypatch): + """The cap used to be read off len(previous_models), which never exceeds four, so any cap above + four was inert. Reading the Router's attempted_retries counter instead lets a cap of five refuse + retries five and six before they reach the deployment.""" + monkeypatch.setattr(litellm, "num_retries_per_request", 5) + router = _always_failing_router(num_retries=6) + + records = await _fail_one_proxy_shaped_request(router, "request-1") + + assert [record["attempted_retries"] for record in records] == [3, 4, 5, 6] + assert ["Max retries per request hit!" in record["exception_string"] for record in records] == [ + False, + False, + True, + True, + ] @pytest.mark.asyncio diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index 835e87aff88..523feb2e54a 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -4061,6 +4061,53 @@ class TestMetadataNoneHandling: assert metadata == {} +_RETRY_CAP_CASES: Final = ( + pytest.param(5, {"attempted_retries": 5}, True, id="cap-above-four-reached"), + pytest.param(5, {"attempted_retries": 4}, False, id="cap-above-four-not-reached"), + pytest.param(0, {"attempted_retries": 0}, False, id="first-attempt-passes-cap-of-zero"), + pytest.param(0, {"attempted_retries": 1}, True, id="cap-of-zero-refuses-first-retry"), + pytest.param(5, {"previous_models": ("a", "b", "c", "d", "e")}, False, id="breadcrumb-count-is-not-the-cap"), + pytest.param(5, None, False, id="metadata-none"), +) + + +def _capped_completion_kwargs(metadata_key: str, metadata: object) -> dict[str, object]: + return { + "model": "openai/gpt-4o-mini", + "messages": [{"role": "user", "content": "hi"}], + "api_key": "sk-fake", + "mock_response": "ok", + metadata_key: metadata, + } + + +@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) +@pytest.mark.parametrize("cap, metadata, refused", _RETRY_CAP_CASES) +def test_num_retries_per_request_reads_attempted_retries_sync(monkeypatch, metadata_key, cap, metadata, refused): + """num_retries_per_request is enforced from the Router's attempted_retries counter in whichever + metadata bucket the call carries, so callers on litellm_metadata and caps above four both work""" + monkeypatch.setattr(litellm, "num_retries_per_request", cap) + kwargs: Final = _capped_completion_kwargs(metadata_key, metadata) + if refused: + with pytest.raises(Exception, match="Max retries per request hit!"): + litellm.completion(**kwargs) + else: + assert litellm.completion(**kwargs).choices[0].message.content == "ok" + + +@pytest.mark.asyncio +@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) +@pytest.mark.parametrize("cap, metadata, refused", _RETRY_CAP_CASES) +async def test_num_retries_per_request_reads_attempted_retries_async(monkeypatch, metadata_key, cap, metadata, refused): + monkeypatch.setattr(litellm, "num_retries_per_request", cap) + kwargs: Final = _capped_completion_kwargs(metadata_key, metadata) + if refused: + with pytest.raises(Exception, match="Max retries per request hit!"): + await litellm.acompletion(**kwargs) + else: + assert (await litellm.acompletion(**kwargs)).choices[0].message.content == "ok" + + class TestValidateAndFixThinkingParam: """Tests for validate_and_fix_thinking_param.""" diff --git a/tests/test_litellm_rust/ocr/test_lifecycle.py b/tests/test_litellm_rust/ocr/test_lifecycle.py index e7ebc5b3018..77d9ef167d0 100644 --- a/tests/test_litellm_rust/ocr/test_lifecycle.py +++ b/tests/test_litellm_rust/ocr/test_lifecycle.py @@ -806,7 +806,7 @@ async def test_shared_call_limits_still_reject_before_reading_ocr_file( monkeypatch.setattr(litellm, "_current_cost", 2) monkeypatch.setattr(litellm, "num_retries_per_request", 1 if limit == "retries" else None) expected: Final = litellm.BudgetExceededError if limit == "budget" else RuntimeError - arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"previous_models": ["earlier"]}} + arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"attempted_retries": 1}} with pytest.raises(expected, match=r"Budget has been exceeded|Max retries per request hit"): await call_aocr(ocr_server, **arguments) if asynchronous else call_ocr(ocr_server, **arguments) assert reads == [] From 566da87771b65665d9d8489898a86db693b5ca06 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 01:20:46 +0000 Subject: [PATCH 007/119] test(router): expect the exact error per retry-cap case and drop explanatory docstrings Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/litellm_core_utils/core_helpers.py | 3 --- tests/test_litellm/test_router.py | 9 +++------ tests/test_litellm/test_utils.py | 2 -- 3 files changed, 3 insertions(+), 11 deletions(-) diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py index ecd9cdac88b..6e76bf9d49e 100644 --- a/litellm/litellm_core_utils/core_helpers.py +++ b/litellm/litellm_core_utils/core_helpers.py @@ -304,9 +304,6 @@ def get_metadata_variable_name_from_kwargs( def max_retries_per_request_hit(kwargs: Mapping[str, object], num_retries_per_request: int | None) -> bool: - """ - Whether the Router retry about to run (``attempted_retries`` >= 1 in the metadata bucket) is past the cap - """ if num_retries_per_request is None: return False metadata: Final = kwargs.get(get_metadata_variable_name_from_kwargs(kwargs)) diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 6150ce287ed..eb29f717a20 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -10746,12 +10746,12 @@ def _always_failing_router(num_retries): ) -async def _fail_one_proxy_shaped_request(router, request_marker): +async def _fail_one_proxy_shaped_request(router, request_marker, expected_error=litellm.InternalServerError): """The proxy hands the router a metadata dict and a proxy_server_request whose body is a shallow copy of the request, so body["metadata"] is the very same dict the router later stamps previous_models onto.""" metadata = {"request_marker": request_marker} - with pytest.raises((litellm.InternalServerError, litellm.APIConnectionError)): + with pytest.raises(expected_error): await router.acompletion( model="broken-group", messages=[{"role": "user", "content": "hi"}], @@ -10808,13 +10808,10 @@ async def test_retry_records_keep_only_the_last_four_attempts(): @pytest.mark.asyncio async def test_num_retries_per_request_stops_retries_at_caps_above_four(monkeypatch): - """The cap used to be read off len(previous_models), which never exceeds four, so any cap above - four was inert. Reading the Router's attempted_retries counter instead lets a cap of five refuse - retries five and six before they reach the deployment.""" monkeypatch.setattr(litellm, "num_retries_per_request", 5) router = _always_failing_router(num_retries=6) - records = await _fail_one_proxy_shaped_request(router, "request-1") + records = await _fail_one_proxy_shaped_request(router, "request-1", expected_error=litellm.APIConnectionError) assert [record["attempted_retries"] for record in records] == [3, 4, 5, 6] assert ["Max retries per request hit!" in record["exception_string"] for record in records] == [ diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index 523feb2e54a..3f3b8ce5343 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -4084,8 +4084,6 @@ def _capped_completion_kwargs(metadata_key: str, metadata: object) -> dict[str, @pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"]) @pytest.mark.parametrize("cap, metadata, refused", _RETRY_CAP_CASES) def test_num_retries_per_request_reads_attempted_retries_sync(monkeypatch, metadata_key, cap, metadata, refused): - """num_retries_per_request is enforced from the Router's attempted_retries counter in whichever - metadata bucket the call carries, so callers on litellm_metadata and caps above four both work""" monkeypatch.setattr(litellm, "num_retries_per_request", cap) kwargs: Final = _capped_completion_kwargs(metadata_key, metadata) if refused: From 6423acc11a1577d70fa7a0aadd7ece09ca21505b Mon Sep 17 00:00:00 2001 From: "berriai-litellm-provider-info-sync[bot]" <328147090+berriai-litellm-provider-info-sync[bot]@users.noreply.github.com> Date: Sun, 13 Sep 2026 04:55:39 +0000 Subject: [PATCH 008/119] chore(prices): sync prices for 5 providers: 278 models, 34 new fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/deepseek-v4-flash-0731: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-vision-exp: fireworks_ai/deepseek-v4-flash-vision-exp: fireworks_ai/accounts/fireworks/models/deepseek-v4-pro: input_cost_per_token, output_cost_per_token, cache_read_input_token_cost, input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/deepseek-v4-pro: input_cost_per_token, output_cost_per_token, cache_read_input_token_cost, input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/deepseek-v4p1-flash: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/deepseek-v4p1-flash: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/glm-5p2: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/glm-5p2: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/glm-5p3: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/glm-5p3-flash: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/gpt-oss-120b: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/gpt-oss-120b: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/kimi-k2p6: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/kimi-k2p6: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/kimi-k2p7-code: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/kimi-k2p7-code: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/kimi-k3: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/kimi-k3: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/minimax-m2p7: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/minimax-m2p7: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/minimax-m3: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/minimax-m3: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/models/muse-glimmer-30b: fireworks_ai/muse-glimmer-30b: fireworks_ai/accounts/fireworks/models/nemotron-3-ultra-nvfp4: fireworks_ai/nemotron-3-ultra-nvfp4: fireworks_ai/accounts/fireworks/models/qwen3-embedding-8b: fireworks_ai/accounts/fireworks/models/qwen3-reranker-8b: input_cost_per_token fireworks_ai/accounts/fireworks/models/qwen3p7-plus: fireworks_ai/qwen3p7-plus: fireworks_ai/accounts/fireworks/models/qwen3p8-max: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/qwen3p8-max: input_cost_per_token_priority, output_cost_per_token_priority, cache_read_input_token_cost_priority fireworks_ai/accounts/fireworks/routers/glm-5p2-fast: fireworks_ai/accounts/fireworks/routers/glm-5p3-fast: input_cost_per_token, output_cost_per_token, cache_read_input_token_cost fireworks_ai/accounts/fireworks/routers/kimi-k3-fast: together_ai/arcee-ai/trinity-mini: input_cost_per_token, output_cost_per_token together_ai/arize-ai/qwen-2-1.5b-instruct: babbage-002: input_cost_per_token_batches, output_cost_per_token_batches chat-latest: chatgpt-image-latest: output_cost_per_token, input_cost_per_image_token, output_cost_per_image_token, input_cost_per_token_batches, output_cost_per_token_batches claude-fable-5: claude-fable-5-1: claude-haiku-4-5: claude-mythos-5: claude-mythos-5-1: claude-opus-4-5: claude-opus-4-6: claude-opus-4-7: claude-opus-4-8: claude-opus-5: claude-sonnet-4-5: claude-sonnet-4-6: claude-sonnet-5: davinci-002: input_cost_per_token_batches, output_cost_per_token_batches deep-research-pro-preview-12-2025: cache_read_input_token_cost together_ai/deepseek-ai/deepseek-coder-33b-instruct: input_cost_per_token, output_cost_per_token together_ai/deepseek-ai/DeepSeek-R1-0528: --- ...odel_prices_and_context_window_backup.json | 968 ++++++++++++++---- model_prices_and_context_window.json | 968 ++++++++++++++---- 2 files changed, 1576 insertions(+), 360 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 2220d0e1fe5..879cf894152 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -11101,12 +11101,15 @@ "babbage-002": { "deprecation_date": "2026-09-28", "input_cost_per_token": 4e-07, + "input_cost_per_token_batches": 2e-07, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 4e-07 + "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "bedrock/*/1-month-commitment/cohere.command-light-text-v14": { "input_cost_per_second": 0.001902, @@ -13171,7 +13174,9 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 0.0001, + "source": "https://developers.openai.com/api/docs/pricing" }, "claude-haiku-4-5-20251001": { "deprecation_date": "2026-10-15", @@ -13219,7 +13224,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-3-7-sonnet-20250219": { "cache_creation_input_token_cost": 3.75e-06, @@ -13378,7 +13384,8 @@ "supports_native_structured_output": true, "supports_tool_choice": true, "supports_vision": true, - "prompt_cache_min_tokens": 1024 + "prompt_cache_min_tokens": 1024, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-5-20250929": { "deprecation_date": "2026-09-29", @@ -13452,7 +13459,7 @@ }, "supports_output_config": true, "prompt_cache_min_tokens": 1024, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-6": { "deprecation_date": "2027-02-17", @@ -13488,7 +13495,8 @@ "prompt_cache_min_tokens": 1024, "provider_specific_entry": { "us": 1.1 - } + }, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-5-20250929-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -13665,7 +13673,8 @@ "supports_tool_choice": true, "supports_vision": true, "supports_output_config": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-6": { "deprecation_date": "2027-02-05", @@ -13702,7 +13711,8 @@ "supports_output_config": true, "supports_max_reasoning_effort": true, "supports_speed": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-6-20260205": { "deprecation_date": "2027-02-05", @@ -13777,7 +13787,8 @@ }, "supports_output_config": true, "supports_speed": true, - "prompt_cache_min_tokens": 2048 + "prompt_cache_min_tokens": 2048, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-7-20260416": { "deprecation_date": "2027-04-16", @@ -13855,7 +13866,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-fable-5-1": { "deprecation_date": "2027-09-01", @@ -13896,7 +13907,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://platform.claude.com/docs/en/models/fable-5-1/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-5": { "deprecation_date": "2027-07-24", @@ -13937,7 +13948,7 @@ "supports_output_config": true, "supports_speed": true, "prompt_cache_min_tokens": 512, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-8": { "deprecation_date": "2027-05-28", @@ -13977,7 +13988,8 @@ }, "supports_output_config": true, "supports_speed": true, - "prompt_cache_min_tokens": 1024 + "prompt_cache_min_tokens": 1024, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-20250514": { "deprecation_date": "2026-06-15", @@ -19246,12 +19258,15 @@ "davinci-002": { "deprecation_date": "2026-09-28", "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 2e-06 + "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "deepgram/base": { "input_cost_per_second": 0.00020833, @@ -22298,15 +22313,18 @@ "supports_vision": false }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro": { - "cache_read_input_token_cost": 1.45e-07, - "input_cost_per_token": 1.74e-06, + "cache_read_input_token_cost": 6e-07, + "cache_read_input_token_cost_priority": 6e-07, + "input_cost_per_token": 1.2e-06, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 384000, "max_tokens": 384000, "mode": "chat", - "output_cost_per_token": 3.48e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token": 1.2e-06, + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22315,14 +22333,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": { "cache_read_input_token_cost": 4.4e-08, + "cache_read_input_token_cost_priority": 5.5e-08, "input_cost_per_token": 1.32e-06, + "input_cost_per_token_priority": 1.65e-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", + "output_cost_per_token_priority": 4.95e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22417,14 +22438,17 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p2": { "cache_read_input_token_cost": 1.4e-07, + "cache_read_input_token_cost_priority": 1.75e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22433,14 +22457,17 @@ }, "fireworks_ai/accounts/fireworks/models/gpt-oss-120b": { "cache_read_input_token_cost": 1.5e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.8e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 131072, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 7.2e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22519,14 +22546,17 @@ }, "fireworks_ai/accounts/fireworks/models/kimi-k2p6": { "cache_read_input_token_cost": 1.6e-07, + "cache_read_input_token_cost_priority": 2.2e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.5e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22535,14 +22565,17 @@ }, "fireworks_ai/accounts/fireworks/models/kimi-k2p7-code": { "cache_read_input_token_cost": 1.9e-07, + "cache_read_input_token_cost_priority": 2.85e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.425e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22668,14 +22701,17 @@ }, "fireworks_ai/accounts/fireworks/models/minimax-m2p7": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 6e-07, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 196608, "max_output_tokens": 196608, "max_tokens": 196608, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22684,14 +22720,17 @@ }, "fireworks_ai/accounts/fireworks/models/minimax-m3": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 9e-08, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 512000, "max_output_tokens": 512000, "max_tokens": 512000, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.8e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22767,15 +22806,18 @@ "supports_vision": false }, "fireworks_ai/deepseek-v4-pro": { - "cache_read_input_token_cost": 1.45e-07, - "input_cost_per_token": 1.74e-06, + "cache_read_input_token_cost": 6e-07, + "cache_read_input_token_cost_priority": 6e-07, + "input_cost_per_token": 1.2e-06, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 384000, "max_tokens": 384000, "mode": "chat", - "output_cost_per_token": 3.48e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token": 1.2e-06, + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22831,14 +22873,17 @@ }, "fireworks_ai/glm-5p2": { "cache_read_input_token_cost": 1.4e-07, + "cache_read_input_token_cost_priority": 1.75e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22847,14 +22892,17 @@ }, "fireworks_ai/gpt-oss-120b": { "cache_read_input_token_cost": 1.5e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.8e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 131072, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 7.2e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22893,14 +22941,17 @@ }, "fireworks_ai/kimi-k2p6": { "cache_read_input_token_cost": 1.6e-07, + "cache_read_input_token_cost_priority": 2.2e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.5e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22925,14 +22976,17 @@ }, "fireworks_ai/kimi-k2p7-code": { "cache_read_input_token_cost": 1.9e-07, + "cache_read_input_token_cost_priority": 2.85e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.425e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22971,14 +23025,17 @@ }, "fireworks_ai/minimax-m2p7": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 6e-07, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 196608, "max_output_tokens": 196608, "max_tokens": 196608, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22987,14 +23044,17 @@ }, "fireworks_ai/minimax-m3": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 9e-08, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 512000, "max_output_tokens": 512000, "max_tokens": 512000, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.8e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -23010,7 +23070,7 @@ "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 1.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -23287,26 +23347,28 @@ "ft:babbage-002": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1.6e-06, - "input_cost_per_token_batches": 2e-07, + "input_cost_per_token_batches": 8e-07, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", "output_cost_per_token": 1.6e-06, - "output_cost_per_token_batches": 2e-07 + "output_cost_per_token_batches": 9e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "ft:davinci-002": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1.2e-05, - "input_cost_per_token_batches": 1e-06, + "input_cost_per_token_batches": 6e-06, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", "output_cost_per_token": 1.2e-05, - "output_cost_per_token_batches": 1e-06 + "output_cost_per_token_batches": 6e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "ft:gpt-3.5-turbo": { "deprecation_date": "2026-10-23", @@ -23319,6 +23381,7 @@ "mode": "chat", "output_cost_per_token": 6e-06, "output_cost_per_token_batches": 3e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_system_messages": true, "supports_tool_choice": true }, @@ -23375,14 +23438,15 @@ "ft:gpt-4o-2024-08-06": { "cache_read_input_token_cost": 1.875e-06, "input_cost_per_token": 3.75e-06, - "input_cost_per_token_batches": 1.875e-06, + "input_cost_per_token_batches": 2.225e-06, "litellm_provider": "openai", "max_input_tokens": 128000, "max_output_tokens": 16384, "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-05, - "output_cost_per_token_batches": 7.5e-06, + "output_cost_per_token_batches": 1.25e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -23420,6 +23484,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-06, "output_cost_per_token_batches": 6e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -23439,6 +23504,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-05, "output_cost_per_token_batches": 6e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23457,6 +23523,7 @@ "mode": "chat", "output_cost_per_token": 3.2e-06, "output_cost_per_token_batches": 1.6e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23476,6 +23543,7 @@ "mode": "chat", "output_cost_per_token": 8e-07, "output_cost_per_token_batches": 4e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23495,6 +23563,7 @@ "mode": "chat", "output_cost_per_token": 1.6e-05, "output_cost_per_token_batches": 8e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -23505,15 +23574,18 @@ "gemini-2.0-flash": { "cache_read_input_token_cost": 2.5e-08, "deprecation_date": "2026-06-01", - "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1e-07, + "input_cost_per_audio_token": 1e-06, + "input_cost_per_character": 3.75e-08, + "input_cost_per_token": 1.5e-07, + "input_cost_per_token_batches": 7.5e-08, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 4e-07, - "source": "https://ai.google.dev/pricing#2_0flash", + "output_cost_per_token": 6e-07, + "output_cost_per_token_batches": 3e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image", @@ -23582,13 +23654,16 @@ "cache_read_input_token_cost": 1.875e-08, "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 7.5e-08, + "input_cost_per_character": 1.875e-08, "input_cost_per_token": 7.5e-08, + "input_cost_per_token_batches": 3.75e-08, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 8192, "mode": "chat", "output_cost_per_token": 3e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", + "output_cost_per_token_batches": 1.5e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image", @@ -23651,6 +23726,8 @@ "gemini-2.5-flash": { "deprecation_date": "2026-10-20", "cache_read_input_token_cost": 3e-08, + "cache_read_input_token_cost_flex": 3e-08, + "cache_read_input_token_cost_priority": 5.4e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "vertex_ai-language-models", @@ -23660,7 +23737,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23693,6 +23770,12 @@ "search_context_size_high": 0.035 }, "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_batches": 1.5e-07, + "input_cost_per_token_flex": 1.5e-07, + "input_cost_per_token_priority": 5.4e-07, + "output_cost_per_token_batches": 1.25e-06, + "output_cost_per_token_flex": 1.25e-06, + "output_cost_per_token_priority": 4.5e-06, "supports_image_size": false }, "gemini-2.5-flash-image": { @@ -23700,6 +23783,9 @@ "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, + "input_cost_per_token_batches": 1.5e-07, + "input_cost_per_token_flex": 1.5e-07, + "input_cost_per_token_priority": 5.4e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 32768, "max_output_tokens": 32768, @@ -23709,8 +23795,10 @@ "output_cost_per_image_token": 3e-05, "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, + "output_cost_per_token_batches": 1.25e-06, + "output_cost_per_token_flex": 1.25e-06, "rpm": 100000, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-flash-image", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23741,10 +23829,19 @@ "supports_image_size": false }, "gemini-3-pro-image": { + "cache_read_input_token_cost": 2e-07, + "cache_read_input_token_cost_above_200k_tokens": 4e-07, + "cache_read_input_token_cost_above_200k_tokens_priority": 7.2e-07, + "cache_read_input_token_cost_flex": 1e-07, + "cache_read_input_token_cost_priority": 3.6e-07, "deprecation_date": "2027-05-28", "input_cost_per_image": 0.0011, "input_cost_per_token": 2e-06, + "input_cost_per_token_above_200k_tokens": 4e-06, + "input_cost_per_token_above_200k_tokens_priority": 7.2e-06, "input_cost_per_token_batches": 1e-06, + "input_cost_per_token_flex": 1e-06, + "input_cost_per_token_priority": 3.6e-06, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 32768, @@ -23753,8 +23850,12 @@ "output_cost_per_image": 0.134, "output_cost_per_image_token": 0.00012, "output_cost_per_token": 1.2e-05, + "output_cost_per_token_above_200k_tokens": 1.8e-05, + "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3-pro-image", + "output_cost_per_token_flex": 6e-06, + "output_cost_per_token_priority": 2.16e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23822,9 +23923,13 @@ "web_search_billing_unit": "per_query" }, "gemini-3.1-flash-image": { + "cache_read_input_token_cost": 5e-08, + "cache_read_input_token_cost_flex": 2.5e-08, "deprecation_date": "2027-05-28", "input_cost_per_image": 0.00056, "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, + "input_cost_per_token_flex": 2.5e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 32768, @@ -23833,7 +23938,9 @@ "output_cost_per_image": 0.0672, "output_cost_per_image_token": 6e-05, "output_cost_per_token": 3e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "output_cost_per_token_batches": 1.5e-06, + "output_cost_per_token_flex": 1.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23900,9 +24007,11 @@ }, "gemini-3.1-flash-lite-image": { "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost_flex": 1.25e-08, "input_cost_per_image": 0.00028, "input_cost_per_token": 2.5e-07, "input_cost_per_token_batches": 1.25e-07, + "input_cost_per_token_flex": 1.25e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 4096, @@ -23912,6 +24021,7 @@ "output_cost_per_image_token": 3e-05, "output_cost_per_token": 1.5e-06, "output_cost_per_token_batches": 7.5e-07, + "output_cost_per_token_flex": 7.5e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -24005,7 +24115,7 @@ "output_cost_per_token_batches": 7.5e-07, "output_cost_per_token_flex": 7.5e-07, "output_cost_per_token_priority": 2.7e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3.1-flash-lite", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24047,7 +24157,7 @@ "deprecation_date": "2027-07-21", "cache_read_input_token_cost": 3e-08, "cache_read_input_token_cost_flex": 1.5e-08, - "cache_read_input_token_cost_priority": 5e-08, + "cache_read_input_token_cost_priority": 5.4e-08, "input_cost_per_token": 3e-07, "input_cost_per_token_batches": 1.5e-07, "input_cost_per_token_flex": 1.5e-07, @@ -24062,7 +24172,7 @@ "output_cost_per_token_batches": 1.25e-06, "output_cost_per_token_flex": 1.25e-06, "output_cost_per_token_priority": 4.5e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24101,6 +24211,7 @@ "google_maps_grounding_cost_per_query": 0.014 }, "deep-research-pro-preview-12-2025": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_image": 0.0011, "input_cost_per_token": 2e-06, "input_cost_per_token_batches": 1e-06, @@ -24113,7 +24224,7 @@ "output_cost_per_image_token": 0.00012, "output_cost_per_token": 1.2e-05, "output_cost_per_token_batches": 6e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24137,6 +24248,8 @@ "gemini-2.5-flash-lite": { "deprecation_date": "2026-10-20", "cache_read_input_token_cost": 1e-08, + "cache_read_input_token_cost_flex": 1e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_audio_token": 3e-07, "input_cost_per_token": 1e-07, "litellm_provider": "vertex_ai-language-models", @@ -24146,7 +24259,7 @@ "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/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24179,6 +24292,12 @@ "search_context_size_high": 0.035 }, "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_batches": 5e-08, + "input_cost_per_token_flex": 5e-08, + "input_cost_per_token_priority": 1.8e-07, + "output_cost_per_token_batches": 2e-07, + "output_cost_per_token_flex": 2e-07, + "output_cost_per_token_priority": 7.2e-07, "supports_image_size": false }, "gemini-2.5-flash-lite-preview-09-2025": { @@ -24330,7 +24449,7 @@ "mode": "realtime", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_token": 2e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/vertex_ai/live" ], @@ -24361,7 +24480,8 @@ "search_context_size_medium": 0.035, "search_context_size_high": 0.035 }, - "gemini_native_audio": true + "gemini_native_audio": true, + "input_cost_per_image_token": 3e-06 }, "gemini/gemini-live-2.5-flash-preview-native-audio-09-2025": { "cache_read_input_token_cost": 7.5e-08, @@ -24461,6 +24581,9 @@ "cache_read_input_token_cost": 1.25e-07, "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, + "cache_read_input_token_cost_above_200k_tokens_priority": 4.5e-07, + "cache_read_input_token_cost_flex": 1.25e-07, + "cache_read_input_token_cost_priority": 2.25e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_above_200k_tokens": 2.5e-06, "litellm_provider": "vertex_ai-language-models", @@ -24470,7 +24593,7 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions" @@ -24500,7 +24623,15 @@ "search_context_size_medium": 0.035, "search_context_size_high": 0.035 }, - "google_maps_grounding_cost_per_query": 0.025 + "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_above_200k_tokens_priority": 4.5e-06, + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "input_cost_per_token_priority": 2.25e-06, + "output_cost_per_token_above_200k_tokens_priority": 2.7e-05, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "output_cost_per_token_priority": 1.8e-05 }, "gemini-3-pro-preview": { "deprecation_date": "2026-03-26", @@ -24575,7 +24706,7 @@ "output_cost_per_token_above_200k_tokens": 1.8e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_image": 0.00012, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24609,13 +24740,16 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "cache_read_input_token_cost_priority": 3.6e-07, "cache_read_input_token_cost_above_200k_tokens_priority": 7.2e-07, + "cache_read_input_token_cost_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_low": 0.014, "search_context_size_medium": 0.014, "search_context_size_high": 0.014 }, "web_search_billing_unit": "per_query", - "google_maps_grounding_cost_per_query": 0.014 + "google_maps_grounding_cost_per_query": 0.014, + "input_cost_per_token_flex": 1e-06, + "output_cost_per_token_flex": 6e-06 }, "gemini-3.1-pro-preview-customtools": { "prompt_cache_min_tokens": 4096, @@ -25127,13 +25261,15 @@ "cache_read_input_token_cost": 1.25e-07, "input_cost_per_audio_token": 7e-07, "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", + "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2e-05, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text" ], @@ -25340,7 +25476,7 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/computer-use", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image" @@ -25381,8 +25517,11 @@ }, "gemini-embedding-2": { "input_cost_per_audio_per_second": 0.00016, + "input_cost_per_audio_token": 6.5e-06, "input_cost_per_image": 0.00012, + "input_cost_per_image_token": 4.5e-07, "input_cost_per_token": 2e-07, + "input_cost_per_token_batches": 1e-07, "input_cost_per_video_per_second": 0.00079, "litellm_provider": "vertex_ai-embedding-models", "max_input_tokens": 8192, @@ -25390,7 +25529,7 @@ "mode": "embedding", "output_cost_per_token": 0, "output_vector_size": 3072, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supports_multimodal": true, "uses_embed_content": true }, @@ -27055,6 +27194,7 @@ }, "gemini-3-flash-preview": { "cache_read_input_token_cost": 5e-08, + "cache_read_input_token_cost_flex": 5e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 5e-07, "litellm_provider": "vertex_ai-language-models", @@ -27064,7 +27204,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 3e-06, "output_cost_per_token": 3e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27102,7 +27242,11 @@ "search_context_size_high": 0.014 }, "web_search_billing_unit": "per_query", - "google_maps_grounding_cost_per_query": 0.014 + "google_maps_grounding_cost_per_query": 0.014, + "input_cost_per_token_batches": 2.5e-07, + "input_cost_per_token_flex": 2.5e-07, + "output_cost_per_token_batches": 1.5e-06, + "output_cost_per_token_flex": 1.5e-06 }, "gemini-omni-flash-preview": { "input_cost_per_audio_token": 1.5e-06, @@ -27115,7 +27259,7 @@ "output_cost_per_reasoning_token": 9e-06, "output_cost_per_token": 9e-06, "output_cost_per_video_token": 1.75e-05, - "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/omni-flash-preview", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions" ], @@ -27148,7 +27292,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 9e-06, "output_cost_per_token": 9e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27211,7 +27355,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27268,7 +27412,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27325,7 +27469,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -28783,6 +28927,7 @@ "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 1.5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_prompt_caching": true, "supports_system_messages": true, @@ -28791,12 +28936,15 @@ "gpt-3.5-turbo-0125": { "deprecation_date": "2026-10-23", "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, "litellm_provider": "openai", "max_input_tokens": 16385, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 1.5e-06, + "output_cost_per_token_batches": 7.5e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -28806,12 +28954,15 @@ "gpt-3.5-turbo-1106": { "deprecation_date": "2026-09-28", "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "openai", "max_input_tokens": 16385, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -28839,7 +28990,8 @@ "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 2e-06 + "output_cost_per_token": 2e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-3.5-turbo-instruct-0914": { "input_cost_per_token": 1.5e-06, @@ -28894,12 +29046,15 @@ "gpt-4-0613": { "deprecation_date": "2026-10-23", "input_cost_per_token": 3e-05, + "input_cost_per_token_batches": 1.5e-05, "litellm_provider": "openai", "max_input_tokens": 8192, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_prompt_caching": true, "supports_system_messages": true, @@ -28940,12 +29095,15 @@ "gpt-4-turbo-2024-04-09": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1e-05, + "input_cost_per_token_batches": 5e-06, "litellm_provider": "openai", "max_input_tokens": 128000, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 3e-05, + "output_cost_per_token_batches": 1.5e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -28989,6 +29147,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29031,6 +29190,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29073,6 +29233,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29115,6 +29276,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29153,6 +29315,7 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_priority": 8e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29190,6 +29353,7 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_priority": 8e-07, "output_cost_per_token_batches": 2e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29226,6 +29390,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_batches": 5e-06, "output_cost_per_token_priority": 1.7e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29248,6 +29413,7 @@ "output_cost_per_token": 1.5e-05, "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_priority": 2.625e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29270,6 +29436,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 1.7e-05, "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29293,6 +29460,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 1.7e-05, "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29367,6 +29535,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29403,6 +29572,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions" ], @@ -29437,6 +29607,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29474,6 +29645,7 @@ "mode": "chat", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29511,6 +29683,7 @@ "mode": "chat", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29572,7 +29745,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": false, - "deprecation_date": "2027-01-20" + "deprecation_date": "2027-01-20", + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini": { "cache_read_input_token_cost": 7.5e-08, @@ -29600,7 +29774,8 @@ "search_context_size_high": 0.025, "search_context_size_low": 0.025, "search_context_size_medium": 0.025 - } + }, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini-2024-07-18": { "cache_read_input_token_cost": 7.5e-08, @@ -29621,6 +29796,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29769,15 +29945,18 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 5e-05, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini-tts": { - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -29909,7 +30088,9 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 0.0001, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-image-1.5": { "cache_read_input_token_cost": 1.25e-06, @@ -29919,7 +30100,10 @@ "mode": "image_generation", "output_cost_per_token": 1e-05, "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations" ], @@ -29934,7 +30118,10 @@ "mode": "image_generation", "output_cost_per_token": 1e-05, "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations" ], @@ -29947,7 +30134,9 @@ "litellm_provider": "openai", "mode": "image_generation", "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -30394,6 +30583,7 @@ "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, + "input_cost_per_token_batches": 6.25e-07, "input_cost_per_token_flex": 6.25e-07, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -30402,6 +30592,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "search_context_cost_per_query": { @@ -30409,6 +30600,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -30438,6 +30630,7 @@ }, "gpt-5.1": { "cache_read_input_token_cost": 1.25e-07, + "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_priority": 2.5e-06, @@ -30478,11 +30671,17 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": false, "supports_minimal_reasoning_effort": false }, "gpt-5.1-2025-11-13": { "cache_read_input_token_cost": 1.25e-07, + "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_priority": 2.5e-06, @@ -30523,6 +30722,11 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": false, "supports_minimal_reasoning_effort": false }, @@ -30574,6 +30778,7 @@ }, "gpt-5.2": { "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_flex": 8.75e-08, "cache_read_input_token_cost_priority": 3.5e-07, "input_cost_per_token": 1.75e-06, "input_cost_per_token_priority": 3.5e-06, @@ -30615,11 +30820,17 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 8.75e-07, + "input_cost_per_token_flex": 8.75e-07, + "output_cost_per_token_batches": 7e-06, + "output_cost_per_token_flex": 7e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, "gpt-5.2-2025-12-11": { "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_flex": 8.75e-08, "cache_read_input_token_cost_priority": 3.5e-07, "input_cost_per_token": 1.75e-06, "input_cost_per_token_priority": 3.5e-06, @@ -30661,6 +30872,11 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 8.75e-07, + "input_cost_per_token_flex": 8.75e-07, + "output_cost_per_token_batches": 7e-06, + "output_cost_per_token_flex": 7e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -30754,17 +30970,20 @@ }, "gpt-5.2-pro": { "input_cost_per_token": 2.1e-05, + "input_cost_per_token_batches": 1.05e-05, "litellm_provider": "openai", "max_input_tokens": 272000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "responses", "output_cost_per_token": 0.000168, + "output_cost_per_token_batches": 8.4e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -30793,17 +31012,20 @@ }, "gpt-5.2-pro-2025-12-11": { "input_cost_per_token": 2.1e-05, + "input_cost_per_token_batches": 1.05e-05, "litellm_provider": "openai", "max_input_tokens": 272000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "responses", "output_cost_per_token": 0.000168, + "output_cost_per_token_batches": 8.4e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -30869,6 +31091,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31005,6 +31228,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31073,6 +31297,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31140,6 +31365,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31203,7 +31429,7 @@ "supports_tool_choice": true, "supports_vision": true, "supports_web_search": true, - "source": "https://developers.openai.com/api/docs/models/gpt-5.6-cyber", + "source": "https://developers.openai.com/api/docs/pricing", "supports_computer_use": true, "supports_parallel_function_calling": true }, @@ -31373,7 +31599,7 @@ "reasoning_effort_levels": [ "medium" ], - "source": "https://developers.openai.com/api/docs/models/chat-latest", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses" @@ -31452,7 +31678,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 5e-06, "output_cost_per_token_above_272k_tokens_flex": 2.25e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.5-2026-04-23": { "cache_read_input_token_cost": 5e-07, @@ -31509,7 +31736,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 5e-06, "output_cost_per_token_above_272k_tokens_flex": 2.25e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.5-pro": { "input_cost_per_token": 3e-05, @@ -31532,6 +31760,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -31580,6 +31809,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -31657,7 +31887,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 2.5e-06, "output_cost_per_token_above_272k_tokens_flex": 1.125e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-2026-03-05": { "cache_read_input_token_cost": 2.5e-07, @@ -31709,7 +31940,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 2.5e-06, "output_cost_per_token_above_272k_tokens_flex": 1.125e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-pro": { "input_cost_per_token": 3e-05, @@ -31758,7 +31990,8 @@ "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 3e-05, - "output_cost_per_token_above_272k_tokens_flex": 0.000135 + "output_cost_per_token_above_272k_tokens_flex": 0.000135, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-pro-2026-03-05": { "input_cost_per_token": 3e-05, @@ -31807,7 +32040,8 @@ "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 3e-05, - "output_cost_per_token_above_272k_tokens_flex": 0.000135 + "output_cost_per_token_above_272k_tokens_flex": 0.000135, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-mini": { "cache_read_input_token_cost": 7.5e-08, @@ -31858,6 +32092,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -31910,6 +32145,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -31959,6 +32195,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -32008,6 +32245,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -32026,6 +32264,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -32068,6 +32307,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -32100,6 +32340,7 @@ "cache_read_input_token_cost_priority": 2.5e-07, "deprecation_date": "2026-12-11", "input_cost_per_token": 1.25e-06, + "input_cost_per_token_batches": 6.25e-07, "input_cost_per_token_flex": 6.25e-07, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -32108,6 +32349,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "search_context_cost_per_query": { @@ -32115,6 +32357,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32439,6 +32682,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses" ], @@ -32469,6 +32713,7 @@ "cache_read_input_token_cost_flex": 1.25e-08, "cache_read_input_token_cost_priority": 4.5e-08, "input_cost_per_token": 2.5e-07, + "input_cost_per_token_batches": 1.25e-07, "input_cost_per_token_flex": 1.25e-07, "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "openai", @@ -32477,6 +32722,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, "search_context_cost_per_query": { @@ -32484,6 +32730,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32517,6 +32764,7 @@ "cache_read_input_token_cost_priority": 4.5e-08, "deprecation_date": "2026-12-11", "input_cost_per_token": 2.5e-07, + "input_cost_per_token_batches": 1.25e-07, "input_cost_per_token_flex": 1.25e-07, "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "openai", @@ -32525,6 +32773,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, "search_context_cost_per_query": { @@ -32532,6 +32781,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32563,6 +32813,7 @@ "cache_read_input_token_cost": 5e-09, "cache_read_input_token_cost_flex": 2.5e-09, "input_cost_per_token": 5e-08, + "input_cost_per_token_batches": 2.5e-08, "input_cost_per_token_flex": 2.5e-08, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -32571,12 +32822,14 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32609,6 +32862,7 @@ "cache_read_input_token_cost_flex": 2.5e-09, "deprecation_date": "2026-12-11", "input_cost_per_token": 5e-08, + "input_cost_per_token_batches": 2.5e-08, "input_cost_per_token_priority": 2.5e-06, "input_cost_per_token_flex": 2.5e-08, "litellm_provider": "openai", @@ -32617,12 +32871,14 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32655,9 +32911,11 @@ "deprecation_date": "2026-10-23", "input_cost_per_image_token": 1e-05, "input_cost_per_token": 5e-06, + "input_cost_per_token_batches": 2.5e-06, "litellm_provider": "openai", "mode": "image_generation", "output_cost_per_image_token": 4e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -32668,9 +32926,11 @@ "deprecation_date": "2026-12-01", "input_cost_per_image_token": 2.5e-06, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "openai", "mode": "image_generation", "output_cost_per_image_token": 8e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -32690,6 +32950,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32722,6 +32983,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32755,6 +33017,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 2.4e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32790,6 +33053,7 @@ "output_cost_per_token": 2.4e-05, "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32825,6 +33089,7 @@ "output_cost_per_token": 2.4e-06, "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32847,8 +33112,10 @@ "gpt-realtime-mini": { "cache_creation_input_audio_token_cost": 3e-07, "cache_read_input_audio_token_cost": 3e-07, + "cache_read_input_token_cost": 6e-08, "deprecation_date": "2027-01-20", "input_cost_per_audio_token": 1e-05, + "input_cost_per_image_token": 8e-07, "input_cost_per_token": 6e-07, "litellm_provider": "openai", "max_input_tokens": 32000, @@ -32857,6 +33124,7 @@ "mode": "realtime", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32890,6 +33158,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -37609,12 +37878,15 @@ "cache_read_input_token_cost": 7.5e-06, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.5e-05, + "input_cost_per_token_batches": 7.5e-06, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -37629,12 +37901,15 @@ "cache_read_input_token_cost": 7.5e-06, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.5e-05, + "input_cost_per_token_batches": 7.5e-06, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -37656,6 +37931,7 @@ "mode": "responses", "output_cost_per_token": 0.0006, "output_cost_per_token_batches": 0.0003, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37689,6 +37965,7 @@ "mode": "responses", "output_cost_per_token": 0.0006, "output_cost_per_token_batches": 0.0003, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37716,6 +37993,7 @@ "cache_read_input_token_cost_flex": 2.5e-07, "cache_read_input_token_cost_priority": 8.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "input_cost_per_token_flex": 1e-06, "input_cost_per_token_priority": 3.5e-06, "litellm_provider": "openai", @@ -37724,6 +38002,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, "search_context_cost_per_query": { @@ -37731,6 +38010,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/chat/completions", @@ -37760,6 +38040,7 @@ "cache_read_input_token_cost_priority": 8.75e-07, "deprecation_date": "2026-12-11", "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "input_cost_per_token_flex": 1e-06, "input_cost_per_token_priority": 3.5e-06, "litellm_provider": "openai", @@ -37768,6 +38049,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, "search_context_cost_per_query": { @@ -37775,6 +38057,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/chat/completions", @@ -37884,12 +38167,15 @@ "cache_read_input_token_cost": 5.5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -37902,12 +38188,15 @@ "cache_read_input_token_cost": 5.5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -37931,6 +38220,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37968,6 +38258,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37991,10 +38282,11 @@ }, "o4-mini": { "cache_read_input_token_cost": 2.75e-07, - "cache_read_input_token_cost_flex": 1.375e-07, + "cache_read_input_token_cost_flex": 1.38e-07, "cache_read_input_token_cost_priority": 5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "input_cost_per_token_flex": 5.5e-07, "input_cost_per_token_priority": 2e-06, "litellm_provider": "openai", @@ -38003,6 +38295,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "output_cost_per_token_flex": 2.2e-06, "output_cost_per_token_priority": 8e-06, "search_context_cost_per_query": { @@ -38010,6 +38303,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -38022,10 +38316,11 @@ }, "o4-mini-2025-04-16": { "cache_read_input_token_cost": 2.75e-07, - "cache_read_input_token_cost_flex": 1.375e-07, + "cache_read_input_token_cost_flex": 1.38e-07, "cache_read_input_token_cost_priority": 5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "input_cost_per_token_flex": 5.5e-07, "input_cost_per_token_priority": 2e-06, "litellm_provider": "openai", @@ -38034,6 +38329,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "output_cost_per_token_flex": 2.2e-06, "output_cost_per_token_priority": 8e-06, "search_context_cost_per_query": { @@ -38041,6 +38337,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -43203,7 +43500,8 @@ "mode": "embedding", "output_cost_per_token": 0.0, "output_cost_per_token_batches": 0.0, - "output_vector_size": 3072 + "output_vector_size": 3072, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-3-small": { "input_cost_per_token": 2e-08, @@ -43214,7 +43512,8 @@ "mode": "embedding", "output_cost_per_token": 0.0, "output_cost_per_token_batches": 0.0, - "output_vector_size": 1536 + "output_vector_size": 1536, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-ada-002": { "input_cost_per_token": 1e-07, @@ -43223,7 +43522,8 @@ "max_tokens": 8191, "mode": "embedding", "output_cost_per_token": 0.0, - "output_vector_size": 1536 + "output_vector_size": 1536, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-ada-002-v2": { "input_cost_per_token": 1e-07, @@ -43394,7 +43694,7 @@ "input_cost_per_token": 1.2e-06, "output_cost_per_token": 1.2e-06, "max_input_tokens": 131072, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen2.5-7B-Instruct-Turbo": { "litellm_provider": "together_ai", @@ -43406,7 +43706,7 @@ "input_cost_per_token": 3e-07, "output_cost_per_token": 3e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-235B-A22B-Instruct-2507-tput": { "deprecation_date": "2026-07-10", @@ -43453,7 +43753,7 @@ "max_input_tokens": 256000, "mode": "chat", "output_cost_per_token": 2e-06, - "source": "https://www.together.ai/models/qwen3-coder-480b-a35b-instruct", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43515,7 +43815,7 @@ }, "mode": "chat", "output_cost_per_token": 1.7e-06, - "source": "https://www.together.ai/models/deepseek-v3-1", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43539,7 +43839,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.04e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43573,6 +43873,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 5.9e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43595,6 +43896,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43606,6 +43908,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 1.8e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43622,7 +43925,7 @@ "input_cost_per_token": 2e-07, "output_cost_per_token": 2e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mistral-Small-24B-Instruct-2501": { "deprecation_date": "2026-04-02", @@ -43634,7 +43937,7 @@ "input_cost_per_token": 1e-07, "output_cost_per_token": 3e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mixtral-8x7B-Instruct-v0.1": { "deprecation_date": "2026-04-16", @@ -43642,6 +43945,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 6e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43668,7 +43972,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://www.together.ai/models/gpt-oss-120b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43682,7 +43986,7 @@ "max_input_tokens": 131072, "mode": "chat", "output_cost_per_token": 2e-07, - "source": "https://www.together.ai/models/gpt-oss-20b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43702,7 +44006,7 @@ "max_input_tokens": 128000, "mode": "chat", "output_cost_per_token": 1.1e-06, - "source": "https://www.together.ai/models/glm-4-5-air", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43718,7 +44022,7 @@ }, "mode": "chat", "output_cost_per_token": 2.2e-06, - "source": "https://www.together.ai/models/glm-4-6", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43735,7 +44039,7 @@ }, "mode": "chat", "output_cost_per_token": 2e-06, - "source": "https://www.together.ai/models/glm-4-7", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43783,7 +44087,7 @@ }, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://www.together.ai/models/qwen3-next-80b-a3b-instruct", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43799,7 +44103,7 @@ }, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://www.together.ai/models/qwen3-next-80b-a3b-thinking", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43813,7 +44117,7 @@ "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 3.6e-06, - "source": "https://www.together.ai/models/qwen3-5-397b-a17b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43828,7 +44132,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43853,7 +44157,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 2.5e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43868,7 +44172,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 3e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { @@ -43879,7 +44183,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 7.5e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { @@ -43889,7 +44193,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 1.28e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { "cache_read_input_token_cost": 2.5e-07, @@ -43899,7 +44203,7 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { @@ -43909,7 +44213,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 1e-07, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { "cache_read_input_token_cost": 3e-08, @@ -43919,7 +44223,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 2.8e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43951,7 +44255,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 3.96e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43976,7 +44280,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 9.7e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -44012,7 +44316,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { @@ -44024,7 +44328,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44045,7 +44349,7 @@ "high", "max" ], - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44063,7 +44367,7 @@ "max_tokens": 512288, "mode": "chat", "output_cost_per_token": 3.6e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44089,7 +44393,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 4.05e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44105,7 +44409,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { @@ -44117,7 +44421,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44134,7 +44438,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44151,7 +44455,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 5e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44164,6 +44468,7 @@ "input_cost_per_character": 1.5e-05, "litellm_provider": "openai", "mode": "audio_speech", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] @@ -44172,6 +44477,7 @@ "input_cost_per_character": 3e-05, "litellm_provider": "openai", "mode": "audio_speech", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] @@ -49497,7 +49803,8 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "source": "https://developers.openai.com/api/docs/pricing" }, "xai/grok-3": { "cache_read_input_token_cost": 2e-07, @@ -52594,10 +52901,11 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 0.0, + "input_cost_per_token": 2e-07, "output_cost_per_token": 0.0, "litellm_provider": "fireworks_ai", - "mode": "rerank" + "mode": "rerank", + "source": "https://api.fireworks.ai/v1/serverless/models" }, "fireworks_ai/accounts/fireworks/models/qwen3-vl-235b-a22b-instruct": { "max_tokens": 262144, @@ -52662,7 +52970,7 @@ "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 1.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -54561,12 +54869,13 @@ }, "gpt-4o-mini-tts-2025-03-20": { "deprecation_date": "2026-07-23", - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -54579,12 +54888,13 @@ ] }, "gpt-4o-mini-tts-2025-12-15": { - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -54599,24 +54909,28 @@ "gpt-4o-mini-transcribe-2025-03-20": { "deprecation_date": "2027-01-20", "input_cost_per_audio_token": 1.25e-06, + "input_cost_per_second": 5e-05, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "mode": "audio_transcription", "output_cost_per_token": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions" ] }, "gpt-4o-mini-transcribe-2025-12-15": { "input_cost_per_audio_token": 1.25e-06, + "input_cost_per_second": 5e-05, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "mode": "audio_transcription", "output_cost_per_token": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions" ] @@ -54635,6 +54949,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -54662,6 +54977,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -54689,6 +55005,7 @@ "mode": "realtime", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -54740,13 +55057,14 @@ "supports_parallel_function_calling": true, "supports_system_messages": true, "supports_tool_choice": true, - "deprecation_date": "2027-01-20" + "deprecation_date": "2027-01-20", + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-realtime-whisper": { - "input_cost_per_second": 0.0002833333333333333, + "input_cost_per_second": 0.000283333333333, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-realtime-whisper", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime", "/v1/realtime/transcription_sessions" @@ -54764,7 +55082,7 @@ "litellm_provider": "openai", "mode": "video_generation", "output_cost_per_video_per_second": 0.1, - "source": "https://platform.openai.com/docs/api-reference/videos", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "image" @@ -54778,7 +55096,7 @@ "litellm_provider": "openai", "mode": "video_generation", "output_cost_per_video_per_second": 0.3, - "source": "https://platform.openai.com/docs/api-reference/videos", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "image" @@ -54803,11 +55121,15 @@ "chatgpt-image-latest": { "cache_read_input_token_cost": 1.25e-06, "deprecation_date": "2026-12-01", - "input_cost_per_image_token": 1e-05, + "input_cost_per_image_token": 8e-06, "input_cost_per_token": 5e-06, + "input_cost_per_token_batches": 2.5e-06, "litellm_provider": "openai", "mode": "image_generation", - "output_cost_per_image_token": 4e-05, + "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -57282,7 +57604,7 @@ "input_cost_per_second": 7.5e-05, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-transcribe", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions", "/v1/realtime/transcription_sessions" @@ -57297,10 +57619,10 @@ "supports_audio_input": true }, "gpt-live-transcribe": { - "input_cost_per_second": 0.0002833333333333333, + "input_cost_per_second": 0.000283333333333, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-live-transcribe", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime", "/v1/realtime/transcription_sessions" @@ -57315,10 +57637,10 @@ "supports_audio_input": true }, "gpt-live-1": { - "input_cost_per_second": 0.0008333333333333334, + "input_cost_per_second": 0.000833333333333, "litellm_provider": "openai", "mode": "realtime", - "source": "https://developers.openai.com/api/docs/models/gpt-live-1", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "audio" @@ -57332,13 +57654,13 @@ "supports_function_calling": true }, "gpt-realtime-translate": { - "input_cost_per_second": 0.0005666666666666667, + "input_cost_per_second": 0.000566666666667, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "max_tokens": 2000, "mode": "realtime", - "source": "https://developers.openai.com/api/docs/models/gpt-realtime-translate", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "audio" ], @@ -57366,7 +57688,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, - "source": "https://platform.claude.com/docs/en/about-claude/models/overview", + "source": "https://platform.claude.com/docs/en/about-claude/pricing", "supports_adaptive_thinking": true, "thinking_always_on": true, "supports_mid_conversation_system": true, @@ -57427,7 +57749,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://platform.claude.com/docs/en/models/mythos-5-1/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-mythos-preview": { "cache_creation_input_token_cost": 1.25e-05, @@ -57779,6 +58101,7 @@ }, "vertex_ai/gemini-3.5-live-translate-preview": { "input_cost_per_audio_token": 3.5e-06, + "input_cost_per_second": 8.83333333333e-05, "input_cost_per_token": 3.5e-06, "litellm_provider": "vertex_ai", "mode": "realtime", @@ -57881,14 +58204,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57897,14 +58223,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4p1-flash": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://fireworks.ai/models/deepseek-ai/deepseek-v4p1-flash", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -57920,26 +58249,29 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_tool_choice": true, "supports_vision": true }, "fireworks_ai/accounts/fireworks/models/kimi-k3": { "cache_read_input_token_cost": 3e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 3e-06, + "input_cost_per_token_priority": 3.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-05, + "output_cost_per_token_priority": 1.875e-05, "reasoning_effort_levels": [ "low", "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57948,14 +58280,17 @@ }, "fireworks_ai/deepseek-v4-flash-0731": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57964,14 +58299,17 @@ }, "fireworks_ai/deepseek-v4p1-flash": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://fireworks.ai/models/deepseek-ai/deepseek-v4p1-flash", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -57987,7 +58325,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_tool_choice": true, "supports_vision": true @@ -58026,19 +58364,22 @@ }, "fireworks_ai/kimi-k3": { "cache_read_input_token_cost": 3e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 3e-06, + "input_cost_per_token_priority": 3.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-05, + "output_cost_per_token_priority": 1.875e-05, "reasoning_effort_levels": [ "low", "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58089,12 +58430,15 @@ }, "fireworks_ai/qwen3p8-max": { "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_priority": 3e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 9e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58110,7 +58454,7 @@ "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58142,7 +58486,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 2.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58158,7 +58502,7 @@ "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58190,7 +58534,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 2.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58199,12 +58543,15 @@ }, "fireworks_ai/accounts/fireworks/models/qwen3p8-max": { "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_priority": 3e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 9e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58220,7 +58567,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58257,7 +58604,7 @@ "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -60955,14 +61302,17 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p3": { "cache_read_input_token_cost": 2.6e-07, + "cache_read_input_token_cost_priority": 3.25e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -60971,13 +61321,16 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p3-flash": { "cache_read_input_token_cost": 3e-08, + "cache_read_input_token_cost_priority": 3.75e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.875e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 5e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -61005,7 +61358,7 @@ "max_output_tokens": 40960, "max_tokens": 40960, "mode": "embedding", - "source": "https://docs.fireworks.ai/serverless/pricing" + "source": "https://api.fireworks.ai/v1/serverless/models" }, "zai/glm-5.2": { "cache_creation_input_token_cost": 0, @@ -61029,7 +61382,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 4.7e-07, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/moonshotai/Kimi-K2.6": { "deprecation_date": "2026-08-19", @@ -61039,7 +61392,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/moonshotai/Kimi-K2.5-fp4": { "input_cost_per_token": 5e-07, @@ -61047,7 +61400,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/MiniMaxAI/MiniMax-M2.7": { "input_cost_per_token": 3e-07, @@ -61056,7 +61409,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 196608, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/zai-org/GLM-5": { "deprecation_date": "2026-06-22", @@ -61065,7 +61418,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 202752, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/zai-org/GLM-5.1": { "deprecation_date": "2026-07-10", @@ -61075,7 +61428,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 202752, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/deepseek-ai/DeepSeek-R1-0528": { "input_cost_per_token": 3e-06, @@ -61083,7 +61436,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 163840, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-Coder-Next-FP8": { "deprecation_date": "2026-05-14", @@ -61092,7 +61445,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-VL-32B-Instruct": { "deprecation_date": "2026-02-25", @@ -61101,7 +61454,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-VL-8B-Instruct": { "deprecation_date": "2026-04-16", @@ -61110,7 +61463,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Ministral-3-14B-Instruct-2512": { "input_cost_per_token": 2e-07, @@ -61118,7 +61471,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/nvidia/NVIDIA-Nemotron-Nano-9B-v2": { "input_cost_per_token": 6e-08, @@ -61126,7 +61479,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 131072, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mistral-7B-Instruct-v0.3": { "input_cost_per_token": 2e-07, @@ -61134,7 +61487,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 32768, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/QwQ-32B": { "deprecation_date": "2025-11-13", @@ -61143,7 +61496,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 131072, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "cerebras/gemma-4-31b": { "input_cost_per_token": 9.9e-07, @@ -65270,5 +65623,260 @@ "supports_tool_choice": false, "supports_response_schema": true, "supports_vision": false + }, + "fireworks_ai/accounts/fireworks/routers/glm-5p3-fast": { + "cache_read_input_token_cost": 3.9e-07, + "input_cost_per_token": 2.1e-06, + "litellm_provider": "fireworks_ai", + "mode": "chat", + "output_cost_per_token": 6.6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models" + }, + "together_ai/arcee-ai/trinity-mini": { + "input_cost_per_token": 4.5e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.5e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/deepseek-coder-33b-instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B": { + "input_cost_per_token": 2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B": { + "input_cost_per_token": 1.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B": { + "input_cost_per_token": 1.6e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.6e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-V4.1-Flash": { + "cache_read_input_token_cost": 6e-09, + "input_cost_per_token": 3e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "vertex_ai/gemini-2.5-flash-native-audio": { + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "realtime", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-2.5-flash-preview-tts": { + "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_audio_token": 1e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.1-flash-live-preview": { + "input_cost_per_audio_token": 3e-06, + "input_cost_per_second": 8.33333333333e-05, + "input_cost_per_token": 7.5e-07, + "litellm_provider": "vertex_ai", + "mode": "realtime", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 4.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.1-flash-tts-preview": { + "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_audio_token": 2e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.5-transcribe": { + "input_cost_per_audio_token": 2e-06, + "input_cost_per_second": 5e-05, + "litellm_provider": "vertex_ai", + "mode": "audio_transcription", + "output_cost_per_token": 1.2e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.5-transcribe-live": { + "input_cost_per_audio_token": 3.5e-06, + "input_cost_per_second": 8.33333333333e-05, + "litellm_provider": "vertex_ai", + "mode": "audio_transcription", + "output_cost_per_token": 2.1e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-omni-1.1-flash": { + "input_cost_per_token": 1.5e-06, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 9e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-robotics-er-2": { + "cache_read_input_token_cost": 1e-07, + "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 5e-06, + "output_cost_per_token_batches": 2.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemma-4-26b-a4b-it": { + "cache_read_input_token_cost": 1.5e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "together_ai/google/gemma-2-27b-it": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "gpt-5.5-cyber": { + "cache_read_input_token_cost": 1.25e-06, + "input_cost_per_token": 1.25e-05, + "litellm_provider": "openai", + "mode": "chat", + "output_cost_per_token": 7.5e-05, + "source": "https://developers.openai.com/api/docs/pricing" + }, + "gpt-rosalind-research": { + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 5e-06, + "litellm_provider": "openai", + "mode": "chat", + "output_cost_per_token": 2.5e-05, + "source": "https://developers.openai.com/api/docs/pricing" + }, + "together_ai/meta-llama/Llama-3-8b-chat-hf": { + "input_cost_per_token": 2e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.1-405B-Instruct": { + "input_cost_per_token": 3.5e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 3.5e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.2-1B-Instruct": { + "input_cost_per_token": 6e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.2-3B-Instruct": { + "input_cost_per_token": 6e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Meta-Llama-3-70B-Instruct-Turbo": { + "input_cost_per_token": 8.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Meta-Llama-3-8B-Instruct": { + "input_cost_per_token": 2e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO": { + "input_cost_per_token": 6e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF": { + "input_cost_per_token": 8.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-1.5B-Instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-72B-Instruct": { + "input_cost_per_token": 9e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 9e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-VL-72B-Instruct": { + "input_cost_per_token": 1.2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-14B-Instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-72B-Instruct": { + "input_cost_per_token": 1.2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-Coder-32B-Instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-VL-72B-Instruct": { + "input_cost_per_token": 1.95e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-06, + "source": "https://api.together.ai/v1/models" } } diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 2220d0e1fe5..879cf894152 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -11101,12 +11101,15 @@ "babbage-002": { "deprecation_date": "2026-09-28", "input_cost_per_token": 4e-07, + "input_cost_per_token_batches": 2e-07, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 4e-07 + "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "bedrock/*/1-month-commitment/cohere.command-light-text-v14": { "input_cost_per_second": 0.001902, @@ -13171,7 +13174,9 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 0.0001, + "source": "https://developers.openai.com/api/docs/pricing" }, "claude-haiku-4-5-20251001": { "deprecation_date": "2026-10-15", @@ -13219,7 +13224,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-3-7-sonnet-20250219": { "cache_creation_input_token_cost": 3.75e-06, @@ -13378,7 +13384,8 @@ "supports_native_structured_output": true, "supports_tool_choice": true, "supports_vision": true, - "prompt_cache_min_tokens": 1024 + "prompt_cache_min_tokens": 1024, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-5-20250929": { "deprecation_date": "2026-09-29", @@ -13452,7 +13459,7 @@ }, "supports_output_config": true, "prompt_cache_min_tokens": 1024, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-6": { "deprecation_date": "2027-02-17", @@ -13488,7 +13495,8 @@ "prompt_cache_min_tokens": 1024, "provider_specific_entry": { "us": 1.1 - } + }, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-5-20250929-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -13665,7 +13673,8 @@ "supports_tool_choice": true, "supports_vision": true, "supports_output_config": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-6": { "deprecation_date": "2027-02-05", @@ -13702,7 +13711,8 @@ "supports_output_config": true, "supports_max_reasoning_effort": true, "supports_speed": true, - "prompt_cache_min_tokens": 4096 + "prompt_cache_min_tokens": 4096, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-6-20260205": { "deprecation_date": "2027-02-05", @@ -13777,7 +13787,8 @@ }, "supports_output_config": true, "supports_speed": true, - "prompt_cache_min_tokens": 2048 + "prompt_cache_min_tokens": 2048, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-7-20260416": { "deprecation_date": "2027-04-16", @@ -13855,7 +13866,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-fable-5-1": { "deprecation_date": "2027-09-01", @@ -13896,7 +13907,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://platform.claude.com/docs/en/models/fable-5-1/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-5": { "deprecation_date": "2027-07-24", @@ -13937,7 +13948,7 @@ "supports_output_config": true, "supports_speed": true, "prompt_cache_min_tokens": 512, - "source": "https://docs.anthropic.com/en/docs/about-claude/models/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-opus-4-8": { "deprecation_date": "2027-05-28", @@ -13977,7 +13988,8 @@ }, "supports_output_config": true, "supports_speed": true, - "prompt_cache_min_tokens": 1024 + "prompt_cache_min_tokens": 1024, + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-sonnet-4-20250514": { "deprecation_date": "2026-06-15", @@ -19246,12 +19258,15 @@ "davinci-002": { "deprecation_date": "2026-09-28", "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 2e-06 + "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "deepgram/base": { "input_cost_per_second": 0.00020833, @@ -22298,15 +22313,18 @@ "supports_vision": false }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro": { - "cache_read_input_token_cost": 1.45e-07, - "input_cost_per_token": 1.74e-06, + "cache_read_input_token_cost": 6e-07, + "cache_read_input_token_cost_priority": 6e-07, + "input_cost_per_token": 1.2e-06, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 384000, "max_tokens": 384000, "mode": "chat", - "output_cost_per_token": 3.48e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token": 1.2e-06, + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22315,14 +22333,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813": { "cache_read_input_token_cost": 4.4e-08, + "cache_read_input_token_cost_priority": 5.5e-08, "input_cost_per_token": 1.32e-06, + "input_cost_per_token_priority": 1.65e-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", + "output_cost_per_token_priority": 4.95e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22417,14 +22438,17 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p2": { "cache_read_input_token_cost": 1.4e-07, + "cache_read_input_token_cost_priority": 1.75e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22433,14 +22457,17 @@ }, "fireworks_ai/accounts/fireworks/models/gpt-oss-120b": { "cache_read_input_token_cost": 1.5e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.8e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 131072, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 7.2e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22519,14 +22546,17 @@ }, "fireworks_ai/accounts/fireworks/models/kimi-k2p6": { "cache_read_input_token_cost": 1.6e-07, + "cache_read_input_token_cost_priority": 2.2e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.5e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22535,14 +22565,17 @@ }, "fireworks_ai/accounts/fireworks/models/kimi-k2p7-code": { "cache_read_input_token_cost": 1.9e-07, + "cache_read_input_token_cost_priority": 2.85e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.425e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22668,14 +22701,17 @@ }, "fireworks_ai/accounts/fireworks/models/minimax-m2p7": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 6e-07, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 196608, "max_output_tokens": 196608, "max_tokens": 196608, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22684,14 +22720,17 @@ }, "fireworks_ai/accounts/fireworks/models/minimax-m3": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 9e-08, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 512000, "max_output_tokens": 512000, "max_tokens": 512000, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.8e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22767,15 +22806,18 @@ "supports_vision": false }, "fireworks_ai/deepseek-v4-pro": { - "cache_read_input_token_cost": 1.45e-07, - "input_cost_per_token": 1.74e-06, + "cache_read_input_token_cost": 6e-07, + "cache_read_input_token_cost_priority": 6e-07, + "input_cost_per_token": 1.2e-06, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 384000, "max_tokens": 384000, "mode": "chat", - "output_cost_per_token": 3.48e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token": 1.2e-06, + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22831,14 +22873,17 @@ }, "fireworks_ai/glm-5p2": { "cache_read_input_token_cost": 1.4e-07, + "cache_read_input_token_cost_priority": 1.75e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22847,14 +22892,17 @@ }, "fireworks_ai/gpt-oss-120b": { "cache_read_input_token_cost": 1.5e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.8e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 131072, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 7.2e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22893,14 +22941,17 @@ }, "fireworks_ai/kimi-k2p6": { "cache_read_input_token_cost": 1.6e-07, + "cache_read_input_token_cost_priority": 2.2e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.5e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22925,14 +22976,17 @@ }, "fireworks_ai/kimi-k2p7-code": { "cache_read_input_token_cost": 1.9e-07, + "cache_read_input_token_cost_priority": 2.85e-07, "input_cost_per_token": 9.5e-07, + "input_cost_per_token_priority": 1.425e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "max_output_tokens": 32768, "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22971,14 +23025,17 @@ }, "fireworks_ai/minimax-m2p7": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 6e-07, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 1.2e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 196608, "max_output_tokens": 196608, "max_tokens": 196608, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.2e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -22987,14 +23044,17 @@ }, "fireworks_ai/minimax-m3": { "cache_read_input_token_cost": 6e-08, + "cache_read_input_token_cost_priority": 9e-08, "input_cost_per_token": 3e-07, + "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 512000, "max_output_tokens": 512000, "max_tokens": 512000, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 1.8e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -23010,7 +23070,7 @@ "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 1.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -23287,26 +23347,28 @@ "ft:babbage-002": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1.6e-06, - "input_cost_per_token_batches": 2e-07, + "input_cost_per_token_batches": 8e-07, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", "output_cost_per_token": 1.6e-06, - "output_cost_per_token_batches": 2e-07 + "output_cost_per_token_batches": 9e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "ft:davinci-002": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1.2e-05, - "input_cost_per_token_batches": 1e-06, + "input_cost_per_token_batches": 6e-06, "litellm_provider": "text-completion-openai", "max_input_tokens": 16384, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", "output_cost_per_token": 1.2e-05, - "output_cost_per_token_batches": 1e-06 + "output_cost_per_token_batches": 6e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "ft:gpt-3.5-turbo": { "deprecation_date": "2026-10-23", @@ -23319,6 +23381,7 @@ "mode": "chat", "output_cost_per_token": 6e-06, "output_cost_per_token_batches": 3e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_system_messages": true, "supports_tool_choice": true }, @@ -23375,14 +23438,15 @@ "ft:gpt-4o-2024-08-06": { "cache_read_input_token_cost": 1.875e-06, "input_cost_per_token": 3.75e-06, - "input_cost_per_token_batches": 1.875e-06, + "input_cost_per_token_batches": 2.225e-06, "litellm_provider": "openai", "max_input_tokens": 128000, "max_output_tokens": 16384, "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-05, - "output_cost_per_token_batches": 7.5e-06, + "output_cost_per_token_batches": 1.25e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -23420,6 +23484,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-06, "output_cost_per_token_batches": 6e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -23439,6 +23504,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-05, "output_cost_per_token_batches": 6e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23457,6 +23523,7 @@ "mode": "chat", "output_cost_per_token": 3.2e-06, "output_cost_per_token_batches": 1.6e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23476,6 +23543,7 @@ "mode": "chat", "output_cost_per_token": 8e-07, "output_cost_per_token_batches": 4e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -23495,6 +23563,7 @@ "mode": "chat", "output_cost_per_token": 1.6e-05, "output_cost_per_token_batches": 8e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -23505,15 +23574,18 @@ "gemini-2.0-flash": { "cache_read_input_token_cost": 2.5e-08, "deprecation_date": "2026-06-01", - "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1e-07, + "input_cost_per_audio_token": 1e-06, + "input_cost_per_character": 3.75e-08, + "input_cost_per_token": 1.5e-07, + "input_cost_per_token_batches": 7.5e-08, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 4e-07, - "source": "https://ai.google.dev/pricing#2_0flash", + "output_cost_per_token": 6e-07, + "output_cost_per_token_batches": 3e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image", @@ -23582,13 +23654,16 @@ "cache_read_input_token_cost": 1.875e-08, "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 7.5e-08, + "input_cost_per_character": 1.875e-08, "input_cost_per_token": 7.5e-08, + "input_cost_per_token_batches": 3.75e-08, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 8192, "mode": "chat", "output_cost_per_token": 3e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", + "output_cost_per_token_batches": 1.5e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image", @@ -23651,6 +23726,8 @@ "gemini-2.5-flash": { "deprecation_date": "2026-10-20", "cache_read_input_token_cost": 3e-08, + "cache_read_input_token_cost_flex": 3e-08, + "cache_read_input_token_cost_priority": 5.4e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, "litellm_provider": "vertex_ai-language-models", @@ -23660,7 +23737,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23693,6 +23770,12 @@ "search_context_size_high": 0.035 }, "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_batches": 1.5e-07, + "input_cost_per_token_flex": 1.5e-07, + "input_cost_per_token_priority": 5.4e-07, + "output_cost_per_token_batches": 1.25e-06, + "output_cost_per_token_flex": 1.25e-06, + "output_cost_per_token_priority": 4.5e-06, "supports_image_size": false }, "gemini-2.5-flash-image": { @@ -23700,6 +23783,9 @@ "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 3e-07, + "input_cost_per_token_batches": 1.5e-07, + "input_cost_per_token_flex": 1.5e-07, + "input_cost_per_token_priority": 5.4e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 32768, "max_output_tokens": 32768, @@ -23709,8 +23795,10 @@ "output_cost_per_image_token": 3e-05, "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, + "output_cost_per_token_batches": 1.25e-06, + "output_cost_per_token_flex": 1.25e-06, "rpm": 100000, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-flash-image", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23741,10 +23829,19 @@ "supports_image_size": false }, "gemini-3-pro-image": { + "cache_read_input_token_cost": 2e-07, + "cache_read_input_token_cost_above_200k_tokens": 4e-07, + "cache_read_input_token_cost_above_200k_tokens_priority": 7.2e-07, + "cache_read_input_token_cost_flex": 1e-07, + "cache_read_input_token_cost_priority": 3.6e-07, "deprecation_date": "2027-05-28", "input_cost_per_image": 0.0011, "input_cost_per_token": 2e-06, + "input_cost_per_token_above_200k_tokens": 4e-06, + "input_cost_per_token_above_200k_tokens_priority": 7.2e-06, "input_cost_per_token_batches": 1e-06, + "input_cost_per_token_flex": 1e-06, + "input_cost_per_token_priority": 3.6e-06, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 32768, @@ -23753,8 +23850,12 @@ "output_cost_per_image": 0.134, "output_cost_per_image_token": 0.00012, "output_cost_per_token": 1.2e-05, + "output_cost_per_token_above_200k_tokens": 1.8e-05, + "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3-pro-image", + "output_cost_per_token_flex": 6e-06, + "output_cost_per_token_priority": 2.16e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23822,9 +23923,13 @@ "web_search_billing_unit": "per_query" }, "gemini-3.1-flash-image": { + "cache_read_input_token_cost": 5e-08, + "cache_read_input_token_cost_flex": 2.5e-08, "deprecation_date": "2027-05-28", "input_cost_per_image": 0.00056, "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, + "input_cost_per_token_flex": 2.5e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 32768, @@ -23833,7 +23938,9 @@ "output_cost_per_image": 0.0672, "output_cost_per_image_token": 6e-05, "output_cost_per_token": 3e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "output_cost_per_token_batches": 1.5e-06, + "output_cost_per_token_flex": 1.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -23900,9 +24007,11 @@ }, "gemini-3.1-flash-lite-image": { "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost_flex": 1.25e-08, "input_cost_per_image": 0.00028, "input_cost_per_token": 2.5e-07, "input_cost_per_token_batches": 1.25e-07, + "input_cost_per_token_flex": 1.25e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 65536, "max_output_tokens": 4096, @@ -23912,6 +24021,7 @@ "output_cost_per_image_token": 3e-05, "output_cost_per_token": 1.5e-06, "output_cost_per_token_batches": 7.5e-07, + "output_cost_per_token_flex": 7.5e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -24005,7 +24115,7 @@ "output_cost_per_token_batches": 7.5e-07, "output_cost_per_token_flex": 7.5e-07, "output_cost_per_token_priority": 2.7e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3.1-flash-lite", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24047,7 +24157,7 @@ "deprecation_date": "2027-07-21", "cache_read_input_token_cost": 3e-08, "cache_read_input_token_cost_flex": 1.5e-08, - "cache_read_input_token_cost_priority": 5e-08, + "cache_read_input_token_cost_priority": 5.4e-08, "input_cost_per_token": 3e-07, "input_cost_per_token_batches": 1.5e-07, "input_cost_per_token_flex": 1.5e-07, @@ -24062,7 +24172,7 @@ "output_cost_per_token_batches": 1.25e-06, "output_cost_per_token_flex": 1.25e-06, "output_cost_per_token_priority": 4.5e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24101,6 +24211,7 @@ "google_maps_grounding_cost_per_query": 0.014 }, "deep-research-pro-preview-12-2025": { + "cache_read_input_token_cost": 2e-07, "input_cost_per_image": 0.0011, "input_cost_per_token": 2e-06, "input_cost_per_token_batches": 1e-06, @@ -24113,7 +24224,7 @@ "output_cost_per_image_token": 0.00012, "output_cost_per_token": 1.2e-05, "output_cost_per_token_batches": 6e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24137,6 +24248,8 @@ "gemini-2.5-flash-lite": { "deprecation_date": "2026-10-20", "cache_read_input_token_cost": 1e-08, + "cache_read_input_token_cost_flex": 1e-08, + "cache_read_input_token_cost_priority": 1.8e-08, "input_cost_per_audio_token": 3e-07, "input_cost_per_token": 1e-07, "litellm_provider": "vertex_ai-language-models", @@ -24146,7 +24259,7 @@ "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/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24179,6 +24292,12 @@ "search_context_size_high": 0.035 }, "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_batches": 5e-08, + "input_cost_per_token_flex": 5e-08, + "input_cost_per_token_priority": 1.8e-07, + "output_cost_per_token_batches": 2e-07, + "output_cost_per_token_flex": 2e-07, + "output_cost_per_token_priority": 7.2e-07, "supports_image_size": false }, "gemini-2.5-flash-lite-preview-09-2025": { @@ -24330,7 +24449,7 @@ "mode": "realtime", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_token": 2e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/vertex_ai/live" ], @@ -24361,7 +24480,8 @@ "search_context_size_medium": 0.035, "search_context_size_high": 0.035 }, - "gemini_native_audio": true + "gemini_native_audio": true, + "input_cost_per_image_token": 3e-06 }, "gemini/gemini-live-2.5-flash-preview-native-audio-09-2025": { "cache_read_input_token_cost": 7.5e-08, @@ -24461,6 +24581,9 @@ "cache_read_input_token_cost": 1.25e-07, "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, + "cache_read_input_token_cost_above_200k_tokens_priority": 4.5e-07, + "cache_read_input_token_cost_flex": 1.25e-07, + "cache_read_input_token_cost_priority": 2.25e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_above_200k_tokens": 2.5e-06, "litellm_provider": "vertex_ai-language-models", @@ -24470,7 +24593,7 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions" @@ -24500,7 +24623,15 @@ "search_context_size_medium": 0.035, "search_context_size_high": 0.035 }, - "google_maps_grounding_cost_per_query": 0.025 + "google_maps_grounding_cost_per_query": 0.025, + "input_cost_per_token_above_200k_tokens_priority": 4.5e-06, + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "input_cost_per_token_priority": 2.25e-06, + "output_cost_per_token_above_200k_tokens_priority": 2.7e-05, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "output_cost_per_token_priority": 1.8e-05 }, "gemini-3-pro-preview": { "deprecation_date": "2026-03-26", @@ -24575,7 +24706,7 @@ "output_cost_per_token_above_200k_tokens": 1.8e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_image": 0.00012, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -24609,13 +24740,16 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "cache_read_input_token_cost_priority": 3.6e-07, "cache_read_input_token_cost_above_200k_tokens_priority": 7.2e-07, + "cache_read_input_token_cost_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_low": 0.014, "search_context_size_medium": 0.014, "search_context_size_high": 0.014 }, "web_search_billing_unit": "per_query", - "google_maps_grounding_cost_per_query": 0.014 + "google_maps_grounding_cost_per_query": 0.014, + "input_cost_per_token_flex": 1e-06, + "output_cost_per_token_flex": 6e-06 }, "gemini-3.1-pro-preview-customtools": { "prompt_cache_min_tokens": 4096, @@ -25127,13 +25261,15 @@ "cache_read_input_token_cost": 1.25e-07, "input_cost_per_audio_token": 7e-07, "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", + "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2e-05, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text" ], @@ -25340,7 +25476,7 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/computer-use", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_modalities": [ "text", "image" @@ -25381,8 +25517,11 @@ }, "gemini-embedding-2": { "input_cost_per_audio_per_second": 0.00016, + "input_cost_per_audio_token": 6.5e-06, "input_cost_per_image": 0.00012, + "input_cost_per_image_token": 4.5e-07, "input_cost_per_token": 2e-07, + "input_cost_per_token_batches": 1e-07, "input_cost_per_video_per_second": 0.00079, "litellm_provider": "vertex_ai-embedding-models", "max_input_tokens": 8192, @@ -25390,7 +25529,7 @@ "mode": "embedding", "output_cost_per_token": 0, "output_vector_size": 3072, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supports_multimodal": true, "uses_embed_content": true }, @@ -27055,6 +27194,7 @@ }, "gemini-3-flash-preview": { "cache_read_input_token_cost": 5e-08, + "cache_read_input_token_cost_flex": 5e-08, "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 5e-07, "litellm_provider": "vertex_ai-language-models", @@ -27064,7 +27204,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 3e-06, "output_cost_per_token": 3e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27102,7 +27242,11 @@ "search_context_size_high": 0.014 }, "web_search_billing_unit": "per_query", - "google_maps_grounding_cost_per_query": 0.014 + "google_maps_grounding_cost_per_query": 0.014, + "input_cost_per_token_batches": 2.5e-07, + "input_cost_per_token_flex": 2.5e-07, + "output_cost_per_token_batches": 1.5e-06, + "output_cost_per_token_flex": 1.5e-06 }, "gemini-omni-flash-preview": { "input_cost_per_audio_token": 1.5e-06, @@ -27115,7 +27259,7 @@ "output_cost_per_reasoning_token": 9e-06, "output_cost_per_token": 9e-06, "output_cost_per_video_token": 1.75e-05, - "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/omni-flash-preview", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions" ], @@ -27148,7 +27292,7 @@ "mode": "chat", "output_cost_per_reasoning_token": 9e-06, "output_cost_per_token": 9e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27211,7 +27355,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27268,7 +27412,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -27325,7 +27469,7 @@ "output_cost_per_token": 3.75e-06, "output_cost_per_token_batches": 1.875e-06, "output_cost_per_token_flex": 1.875e-06, - "source": "https://ai.google.dev/gemini-api/docs/pricing", + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/completions", @@ -28783,6 +28927,7 @@ "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 1.5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_prompt_caching": true, "supports_system_messages": true, @@ -28791,12 +28936,15 @@ "gpt-3.5-turbo-0125": { "deprecation_date": "2026-10-23", "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, "litellm_provider": "openai", "max_input_tokens": 16385, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 1.5e-06, + "output_cost_per_token_batches": 7.5e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -28806,12 +28954,15 @@ "gpt-3.5-turbo-1106": { "deprecation_date": "2026-09-28", "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "openai", "max_input_tokens": 16385, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -28839,7 +28990,8 @@ "max_output_tokens": 4096, "max_tokens": 4096, "mode": "completion", - "output_cost_per_token": 2e-06 + "output_cost_per_token": 2e-06, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-3.5-turbo-instruct-0914": { "input_cost_per_token": 1.5e-06, @@ -28894,12 +29046,15 @@ "gpt-4-0613": { "deprecation_date": "2026-10-23", "input_cost_per_token": 3e-05, + "input_cost_per_token_batches": 1.5e-05, "litellm_provider": "openai", "max_input_tokens": 8192, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_prompt_caching": true, "supports_system_messages": true, @@ -28940,12 +29095,15 @@ "gpt-4-turbo-2024-04-09": { "deprecation_date": "2026-10-23", "input_cost_per_token": 1e-05, + "input_cost_per_token_batches": 5e-06, "litellm_provider": "openai", "max_input_tokens": 128000, "max_output_tokens": 4096, "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 3e-05, + "output_cost_per_token_batches": 1.5e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -28989,6 +29147,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29031,6 +29190,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29073,6 +29233,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29115,6 +29276,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29153,6 +29315,7 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_priority": 8e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29190,6 +29353,7 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_priority": 8e-07, "output_cost_per_token_batches": 2e-07, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -29226,6 +29390,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_batches": 5e-06, "output_cost_per_token_priority": 1.7e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29248,6 +29413,7 @@ "output_cost_per_token": 1.5e-05, "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_priority": 2.625e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29270,6 +29436,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 1.7e-05, "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29293,6 +29460,7 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 1.7e-05, "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29367,6 +29535,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29403,6 +29572,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions" ], @@ -29437,6 +29607,7 @@ "mode": "chat", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29474,6 +29645,7 @@ "mode": "chat", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29511,6 +29683,7 @@ "mode": "chat", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses", @@ -29572,7 +29745,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": false, - "deprecation_date": "2027-01-20" + "deprecation_date": "2027-01-20", + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini": { "cache_read_input_token_cost": 7.5e-08, @@ -29600,7 +29774,8 @@ "search_context_size_high": 0.025, "search_context_size_low": 0.025, "search_context_size_medium": 0.025 - } + }, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini-2024-07-18": { "cache_read_input_token_cost": 7.5e-08, @@ -29621,6 +29796,7 @@ "search_context_size_low": 0.025, "search_context_size_medium": 0.025 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -29769,15 +29945,18 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 5e-05, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-4o-mini-tts": { - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -29909,7 +30088,9 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "input_cost_per_second": 0.0001, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-image-1.5": { "cache_read_input_token_cost": 1.25e-06, @@ -29919,7 +30100,10 @@ "mode": "image_generation", "output_cost_per_token": 1e-05, "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations" ], @@ -29934,7 +30118,10 @@ "mode": "image_generation", "output_cost_per_token": 1e-05, "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations" ], @@ -29947,7 +30134,9 @@ "litellm_provider": "openai", "mode": "image_generation", "input_cost_per_image_token": 8e-06, + "input_cost_per_token_batches": 2.5e-06, "output_cost_per_image_token": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -30394,6 +30583,7 @@ "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, + "input_cost_per_token_batches": 6.25e-07, "input_cost_per_token_flex": 6.25e-07, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -30402,6 +30592,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "search_context_cost_per_query": { @@ -30409,6 +30600,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -30438,6 +30630,7 @@ }, "gpt-5.1": { "cache_read_input_token_cost": 1.25e-07, + "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_priority": 2.5e-06, @@ -30478,11 +30671,17 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": false, "supports_minimal_reasoning_effort": false }, "gpt-5.1-2025-11-13": { "cache_read_input_token_cost": 1.25e-07, + "cache_read_input_token_cost_flex": 6.25e-08, "cache_read_input_token_cost_priority": 2.5e-07, "input_cost_per_token": 1.25e-06, "input_cost_per_token_priority": 2.5e-06, @@ -30523,6 +30722,11 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 6.25e-07, + "input_cost_per_token_flex": 6.25e-07, + "output_cost_per_token_batches": 5e-06, + "output_cost_per_token_flex": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": false, "supports_minimal_reasoning_effort": false }, @@ -30574,6 +30778,7 @@ }, "gpt-5.2": { "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_flex": 8.75e-08, "cache_read_input_token_cost_priority": 3.5e-07, "input_cost_per_token": 1.75e-06, "input_cost_per_token_priority": 3.5e-06, @@ -30615,11 +30820,17 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 8.75e-07, + "input_cost_per_token_flex": 8.75e-07, + "output_cost_per_token_batches": 7e-06, + "output_cost_per_token_flex": 7e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, "gpt-5.2-2025-12-11": { "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_flex": 8.75e-08, "cache_read_input_token_cost_priority": 3.5e-07, "input_cost_per_token": 1.75e-06, "input_cost_per_token_priority": 3.5e-06, @@ -30661,6 +30872,11 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "input_cost_per_token_batches": 8.75e-07, + "input_cost_per_token_flex": 8.75e-07, + "output_cost_per_token_batches": 7e-06, + "output_cost_per_token_flex": 7e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -30754,17 +30970,20 @@ }, "gpt-5.2-pro": { "input_cost_per_token": 2.1e-05, + "input_cost_per_token_batches": 1.05e-05, "litellm_provider": "openai", "max_input_tokens": 272000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "responses", "output_cost_per_token": 0.000168, + "output_cost_per_token_batches": 8.4e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -30793,17 +31012,20 @@ }, "gpt-5.2-pro-2025-12-11": { "input_cost_per_token": 2.1e-05, + "input_cost_per_token_batches": 1.05e-05, "litellm_provider": "openai", "max_input_tokens": 272000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "responses", "output_cost_per_token": 0.000168, + "output_cost_per_token_batches": 8.4e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -30869,6 +31091,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31005,6 +31228,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31073,6 +31297,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31140,6 +31365,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -31203,7 +31429,7 @@ "supports_tool_choice": true, "supports_vision": true, "supports_web_search": true, - "source": "https://developers.openai.com/api/docs/models/gpt-5.6-cyber", + "source": "https://developers.openai.com/api/docs/pricing", "supports_computer_use": true, "supports_parallel_function_calling": true }, @@ -31373,7 +31599,7 @@ "reasoning_effort_levels": [ "medium" ], - "source": "https://developers.openai.com/api/docs/models/chat-latest", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/responses" @@ -31452,7 +31678,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 5e-06, "output_cost_per_token_above_272k_tokens_flex": 2.25e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.5-2026-04-23": { "cache_read_input_token_cost": 5e-07, @@ -31509,7 +31736,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 5e-06, "output_cost_per_token_above_272k_tokens_flex": 2.25e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.5-pro": { "input_cost_per_token": 3e-05, @@ -31532,6 +31760,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -31580,6 +31809,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -31657,7 +31887,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 2.5e-06, "output_cost_per_token_above_272k_tokens_flex": 1.125e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-2026-03-05": { "cache_read_input_token_cost": 2.5e-07, @@ -31709,7 +31940,8 @@ "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 2.5e-06, "output_cost_per_token_above_272k_tokens_flex": 1.125e-05, - "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07 + "cache_read_input_token_cost_above_272k_tokens_flex": 2.5e-07, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-pro": { "input_cost_per_token": 3e-05, @@ -31758,7 +31990,8 @@ "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 3e-05, - "output_cost_per_token_above_272k_tokens_flex": 0.000135 + "output_cost_per_token_above_272k_tokens_flex": 0.000135, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-pro-2026-03-05": { "input_cost_per_token": 3e-05, @@ -31807,7 +32040,8 @@ "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false, "input_cost_per_token_above_272k_tokens_flex": 3e-05, - "output_cost_per_token_above_272k_tokens_flex": 0.000135 + "output_cost_per_token_above_272k_tokens_flex": 0.000135, + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-5.4-mini": { "cache_read_input_token_cost": 7.5e-08, @@ -31858,6 +32092,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -31910,6 +32145,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -31959,6 +32195,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -32008,6 +32245,7 @@ "supports_web_search": true, "supports_none_reasoning_effort": true, "default_reasoning_effort": "none", + "source": "https://developers.openai.com/api/docs/pricing", "supports_xhigh_reasoning_effort": true, "supports_minimal_reasoning_effort": false }, @@ -32026,6 +32264,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -32068,6 +32307,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/batch", "/v1/responses" @@ -32100,6 +32340,7 @@ "cache_read_input_token_cost_priority": 2.5e-07, "deprecation_date": "2026-12-11", "input_cost_per_token": 1.25e-06, + "input_cost_per_token_batches": 6.25e-07, "input_cost_per_token_flex": 6.25e-07, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -32108,6 +32349,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "search_context_cost_per_query": { @@ -32115,6 +32357,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32439,6 +32682,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses" ], @@ -32469,6 +32713,7 @@ "cache_read_input_token_cost_flex": 1.25e-08, "cache_read_input_token_cost_priority": 4.5e-08, "input_cost_per_token": 2.5e-07, + "input_cost_per_token_batches": 1.25e-07, "input_cost_per_token_flex": 1.25e-07, "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "openai", @@ -32477,6 +32722,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, "search_context_cost_per_query": { @@ -32484,6 +32730,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32517,6 +32764,7 @@ "cache_read_input_token_cost_priority": 4.5e-08, "deprecation_date": "2026-12-11", "input_cost_per_token": 2.5e-07, + "input_cost_per_token_batches": 1.25e-07, "input_cost_per_token_flex": 1.25e-07, "input_cost_per_token_priority": 4.5e-07, "litellm_provider": "openai", @@ -32525,6 +32773,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2e-06, + "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, "search_context_cost_per_query": { @@ -32532,6 +32781,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32563,6 +32813,7 @@ "cache_read_input_token_cost": 5e-09, "cache_read_input_token_cost_flex": 2.5e-09, "input_cost_per_token": 5e-08, + "input_cost_per_token_batches": 2.5e-08, "input_cost_per_token_flex": 2.5e-08, "input_cost_per_token_priority": 2.5e-06, "litellm_provider": "openai", @@ -32571,12 +32822,14 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32609,6 +32862,7 @@ "cache_read_input_token_cost_flex": 2.5e-09, "deprecation_date": "2026-12-11", "input_cost_per_token": 5e-08, + "input_cost_per_token_batches": 2.5e-08, "input_cost_per_token_priority": 2.5e-06, "input_cost_per_token_flex": 2.5e-08, "litellm_provider": "openai", @@ -32617,12 +32871,14 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4e-07, + "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -32655,9 +32911,11 @@ "deprecation_date": "2026-10-23", "input_cost_per_image_token": 1e-05, "input_cost_per_token": 5e-06, + "input_cost_per_token_batches": 2.5e-06, "litellm_provider": "openai", "mode": "image_generation", "output_cost_per_image_token": 4e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -32668,9 +32926,11 @@ "deprecation_date": "2026-12-01", "input_cost_per_image_token": 2.5e-06, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "openai", "mode": "image_generation", "output_cost_per_image_token": 8e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -32690,6 +32950,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32722,6 +32983,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32755,6 +33017,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 2.4e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32790,6 +33053,7 @@ "output_cost_per_token": 2.4e-05, "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32825,6 +33089,7 @@ "output_cost_per_token": 2.4e-06, "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32847,8 +33112,10 @@ "gpt-realtime-mini": { "cache_creation_input_audio_token_cost": 3e-07, "cache_read_input_audio_token_cost": 3e-07, + "cache_read_input_token_cost": 6e-08, "deprecation_date": "2027-01-20", "input_cost_per_audio_token": 1e-05, + "input_cost_per_image_token": 8e-07, "input_cost_per_token": 6e-07, "litellm_provider": "openai", "max_input_tokens": 32000, @@ -32857,6 +33124,7 @@ "mode": "realtime", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -32890,6 +33158,7 @@ "mode": "realtime", "output_cost_per_audio_token": 6.4e-05, "output_cost_per_token": 1.6e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -37609,12 +37878,15 @@ "cache_read_input_token_cost": 7.5e-06, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.5e-05, + "input_cost_per_token_batches": 7.5e-06, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -37629,12 +37901,15 @@ "cache_read_input_token_cost": 7.5e-06, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.5e-05, + "input_cost_per_token_batches": 7.5e-06, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 6e-05, + "output_cost_per_token_batches": 3e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -37656,6 +37931,7 @@ "mode": "responses", "output_cost_per_token": 0.0006, "output_cost_per_token_batches": 0.0003, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37689,6 +37965,7 @@ "mode": "responses", "output_cost_per_token": 0.0006, "output_cost_per_token_batches": 0.0003, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37716,6 +37993,7 @@ "cache_read_input_token_cost_flex": 2.5e-07, "cache_read_input_token_cost_priority": 8.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "input_cost_per_token_flex": 1e-06, "input_cost_per_token_priority": 3.5e-06, "litellm_provider": "openai", @@ -37724,6 +38002,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, "search_context_cost_per_query": { @@ -37731,6 +38010,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/chat/completions", @@ -37760,6 +38040,7 @@ "cache_read_input_token_cost_priority": 8.75e-07, "deprecation_date": "2026-12-11", "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "input_cost_per_token_flex": 1e-06, "input_cost_per_token_priority": 3.5e-06, "litellm_provider": "openai", @@ -37768,6 +38049,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, "search_context_cost_per_query": { @@ -37775,6 +38057,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/chat/completions", @@ -37884,12 +38167,15 @@ "cache_read_input_token_cost": 5.5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -37902,12 +38188,15 @@ "cache_read_input_token_cost": 5.5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "openai", "max_input_tokens": 200000, "max_output_tokens": 100000, "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_prompt_caching": true, @@ -37931,6 +38220,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37968,6 +38258,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/responses", "/v1/batch" @@ -37991,10 +38282,11 @@ }, "o4-mini": { "cache_read_input_token_cost": 2.75e-07, - "cache_read_input_token_cost_flex": 1.375e-07, + "cache_read_input_token_cost_flex": 1.38e-07, "cache_read_input_token_cost_priority": 5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "input_cost_per_token_flex": 5.5e-07, "input_cost_per_token_priority": 2e-06, "litellm_provider": "openai", @@ -38003,6 +38295,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "output_cost_per_token_flex": 2.2e-06, "output_cost_per_token_priority": 8e-06, "search_context_cost_per_query": { @@ -38010,6 +38303,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -38022,10 +38316,11 @@ }, "o4-mini-2025-04-16": { "cache_read_input_token_cost": 2.75e-07, - "cache_read_input_token_cost_flex": 1.375e-07, + "cache_read_input_token_cost_flex": 1.38e-07, "cache_read_input_token_cost_priority": 5e-07, "deprecation_date": "2026-10-23", "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "input_cost_per_token_flex": 5.5e-07, "input_cost_per_token_priority": 2e-06, "litellm_provider": "openai", @@ -38034,6 +38329,7 @@ "max_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "output_cost_per_token_flex": 2.2e-06, "output_cost_per_token_priority": 8e-06, "search_context_cost_per_query": { @@ -38041,6 +38337,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, @@ -43203,7 +43500,8 @@ "mode": "embedding", "output_cost_per_token": 0.0, "output_cost_per_token_batches": 0.0, - "output_vector_size": 3072 + "output_vector_size": 3072, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-3-small": { "input_cost_per_token": 2e-08, @@ -43214,7 +43512,8 @@ "mode": "embedding", "output_cost_per_token": 0.0, "output_cost_per_token_batches": 0.0, - "output_vector_size": 1536 + "output_vector_size": 1536, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-ada-002": { "input_cost_per_token": 1e-07, @@ -43223,7 +43522,8 @@ "max_tokens": 8191, "mode": "embedding", "output_cost_per_token": 0.0, - "output_vector_size": 1536 + "output_vector_size": 1536, + "source": "https://developers.openai.com/api/docs/pricing" }, "text-embedding-ada-002-v2": { "input_cost_per_token": 1e-07, @@ -43394,7 +43694,7 @@ "input_cost_per_token": 1.2e-06, "output_cost_per_token": 1.2e-06, "max_input_tokens": 131072, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen2.5-7B-Instruct-Turbo": { "litellm_provider": "together_ai", @@ -43406,7 +43706,7 @@ "input_cost_per_token": 3e-07, "output_cost_per_token": 3e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-235B-A22B-Instruct-2507-tput": { "deprecation_date": "2026-07-10", @@ -43453,7 +43753,7 @@ "max_input_tokens": 256000, "mode": "chat", "output_cost_per_token": 2e-06, - "source": "https://www.together.ai/models/qwen3-coder-480b-a35b-instruct", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43515,7 +43815,7 @@ }, "mode": "chat", "output_cost_per_token": 1.7e-06, - "source": "https://www.together.ai/models/deepseek-v3-1", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43539,7 +43839,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.04e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43573,6 +43873,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 5.9e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43595,6 +43896,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43606,6 +43908,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 1.8e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43622,7 +43925,7 @@ "input_cost_per_token": 2e-07, "output_cost_per_token": 2e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mistral-Small-24B-Instruct-2501": { "deprecation_date": "2026-04-02", @@ -43634,7 +43937,7 @@ "input_cost_per_token": 1e-07, "output_cost_per_token": 3e-07, "max_input_tokens": 32768, - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mixtral-8x7B-Instruct-v0.1": { "deprecation_date": "2026-04-16", @@ -43642,6 +43945,7 @@ "litellm_provider": "together_ai", "mode": "chat", "output_cost_per_token": 6e-07, + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43668,7 +43972,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://www.together.ai/models/gpt-oss-120b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43682,7 +43986,7 @@ "max_input_tokens": 131072, "mode": "chat", "output_cost_per_token": 2e-07, - "source": "https://www.together.ai/models/gpt-oss-20b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43702,7 +44006,7 @@ "max_input_tokens": 128000, "mode": "chat", "output_cost_per_token": 1.1e-06, - "source": "https://www.together.ai/models/glm-4-5-air", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43718,7 +44022,7 @@ }, "mode": "chat", "output_cost_per_token": 2.2e-06, - "source": "https://www.together.ai/models/glm-4-6", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43735,7 +44039,7 @@ }, "mode": "chat", "output_cost_per_token": 2e-06, - "source": "https://www.together.ai/models/glm-4-7", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43783,7 +44087,7 @@ }, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://www.together.ai/models/qwen3-next-80b-a3b-instruct", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43799,7 +44103,7 @@ }, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://www.together.ai/models/qwen3-next-80b-a3b-thinking", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -43813,7 +44117,7 @@ "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 3.6e-06, - "source": "https://www.together.ai/models/qwen3-5-397b-a17b", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43828,7 +44132,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43853,7 +44157,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 2.5e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_reasoning": true, @@ -43868,7 +44172,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 3e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_reasoning": true }, "together_ai/Qwen/Qwen3.7-Max": { @@ -43879,7 +44183,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 7.5e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/Qwen/Qwen3.7-Plus": { @@ -43889,7 +44193,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 1.28e-06, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3.8-2.4T-A95B": { "cache_read_input_token_cost": 2.5e-07, @@ -43899,7 +44203,7 @@ "max_tokens": 1010000, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/arize-ai/qwen-2-1.5b-instruct": { @@ -43909,7 +44213,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 1e-07, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/deepseek-ai/DeepSeek-V4-Flash-0731": { "cache_read_input_token_cost": 3e-08, @@ -43919,7 +44223,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 2.8e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43951,7 +44255,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 3.96e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -43976,7 +44280,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 9.7e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_response_schema": true, @@ -44012,7 +44316,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/moonshotai/Kimi-K2.7-Code": { @@ -44024,7 +44328,7 @@ "max_tokens": 262144, "mode": "chat", "output_cost_per_token": 4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44045,7 +44349,7 @@ "high", "max" ], - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44063,7 +44367,7 @@ "max_tokens": 512288, "mode": "chat", "output_cost_per_token": 3.6e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44089,7 +44393,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 4.05e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44105,7 +44409,7 @@ "max_tokens": 524288, "mode": "chat", "output_cost_per_token": 1.2e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_prompt_caching": true }, "together_ai/zai-org/GLM-5.2": { @@ -44117,7 +44421,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44134,7 +44438,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44151,7 +44455,7 @@ "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 5e-07, - "source": "https://docs.together.ai/docs/serverless-models", + "source": "https://api.together.ai/v1/models", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -44164,6 +44468,7 @@ "input_cost_per_character": 1.5e-05, "litellm_provider": "openai", "mode": "audio_speech", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] @@ -44172,6 +44477,7 @@ "input_cost_per_character": 3e-05, "litellm_provider": "openai", "mode": "audio_speech", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] @@ -49497,7 +49803,8 @@ "supported_endpoints": [ "/v1/audio/transcriptions" ], - "deprecation_date": "2027-02-26" + "deprecation_date": "2027-02-26", + "source": "https://developers.openai.com/api/docs/pricing" }, "xai/grok-3": { "cache_read_input_token_cost": 2e-07, @@ -52594,10 +52901,11 @@ "max_tokens": 40960, "max_input_tokens": 40960, "max_output_tokens": 40960, - "input_cost_per_token": 0.0, + "input_cost_per_token": 2e-07, "output_cost_per_token": 0.0, "litellm_provider": "fireworks_ai", - "mode": "rerank" + "mode": "rerank", + "source": "https://api.fireworks.ai/v1/serverless/models" }, "fireworks_ai/accounts/fireworks/models/qwen3-vl-235b-a22b-instruct": { "max_tokens": 262144, @@ -52662,7 +52970,7 @@ "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 1.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -54561,12 +54869,13 @@ }, "gpt-4o-mini-tts-2025-03-20": { "deprecation_date": "2026-07-23", - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -54579,12 +54888,13 @@ ] }, "gpt-4o-mini-tts-2025-12-15": { - "input_cost_per_token": 2.5e-06, + "input_cost_per_token": 6e-07, "litellm_provider": "openai", "mode": "audio_speech", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_second": 0.00025, "output_cost_per_token": 1e-05, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -54599,24 +54909,28 @@ "gpt-4o-mini-transcribe-2025-03-20": { "deprecation_date": "2027-01-20", "input_cost_per_audio_token": 1.25e-06, + "input_cost_per_second": 5e-05, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "mode": "audio_transcription", "output_cost_per_token": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions" ] }, "gpt-4o-mini-transcribe-2025-12-15": { "input_cost_per_audio_token": 1.25e-06, + "input_cost_per_second": 5e-05, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "mode": "audio_transcription", "output_cost_per_token": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions" ] @@ -54635,6 +54949,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -54662,6 +54977,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "source": "https://developers.openai.com/api/docs/pricing", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, @@ -54689,6 +55005,7 @@ "mode": "realtime", "output_cost_per_audio_token": 2e-05, "output_cost_per_token": 2.4e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -54740,13 +55057,14 @@ "supports_parallel_function_calling": true, "supports_system_messages": true, "supports_tool_choice": true, - "deprecation_date": "2027-01-20" + "deprecation_date": "2027-01-20", + "source": "https://developers.openai.com/api/docs/pricing" }, "gpt-realtime-whisper": { - "input_cost_per_second": 0.0002833333333333333, + "input_cost_per_second": 0.000283333333333, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-realtime-whisper", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime", "/v1/realtime/transcription_sessions" @@ -54764,7 +55082,7 @@ "litellm_provider": "openai", "mode": "video_generation", "output_cost_per_video_per_second": 0.1, - "source": "https://platform.openai.com/docs/api-reference/videos", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "image" @@ -54778,7 +55096,7 @@ "litellm_provider": "openai", "mode": "video_generation", "output_cost_per_video_per_second": 0.3, - "source": "https://platform.openai.com/docs/api-reference/videos", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "image" @@ -54803,11 +55121,15 @@ "chatgpt-image-latest": { "cache_read_input_token_cost": 1.25e-06, "deprecation_date": "2026-12-01", - "input_cost_per_image_token": 1e-05, + "input_cost_per_image_token": 8e-06, "input_cost_per_token": 5e-06, + "input_cost_per_token_batches": 2.5e-06, "litellm_provider": "openai", "mode": "image_generation", - "output_cost_per_image_token": 4e-05, + "output_cost_per_image_token": 3.2e-05, + "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/images/generations", "/v1/images/edits" @@ -57282,7 +57604,7 @@ "input_cost_per_second": 7.5e-05, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-transcribe", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/audio/transcriptions", "/v1/realtime/transcription_sessions" @@ -57297,10 +57619,10 @@ "supports_audio_input": true }, "gpt-live-transcribe": { - "input_cost_per_second": 0.0002833333333333333, + "input_cost_per_second": 0.000283333333333, "litellm_provider": "openai", "mode": "audio_transcription", - "source": "https://developers.openai.com/api/docs/models/gpt-live-transcribe", + "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/realtime", "/v1/realtime/transcription_sessions" @@ -57315,10 +57637,10 @@ "supports_audio_input": true }, "gpt-live-1": { - "input_cost_per_second": 0.0008333333333333334, + "input_cost_per_second": 0.000833333333333, "litellm_provider": "openai", "mode": "realtime", - "source": "https://developers.openai.com/api/docs/models/gpt-live-1", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "text", "audio" @@ -57332,13 +57654,13 @@ "supports_function_calling": true }, "gpt-realtime-translate": { - "input_cost_per_second": 0.0005666666666666667, + "input_cost_per_second": 0.000566666666667, "litellm_provider": "openai", "max_input_tokens": 16000, "max_output_tokens": 2000, "max_tokens": 2000, "mode": "realtime", - "source": "https://developers.openai.com/api/docs/models/gpt-realtime-translate", + "source": "https://developers.openai.com/api/docs/pricing", "supported_modalities": [ "audio" ], @@ -57366,7 +57688,7 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, - "source": "https://platform.claude.com/docs/en/about-claude/models/overview", + "source": "https://platform.claude.com/docs/en/about-claude/pricing", "supports_adaptive_thinking": true, "thinking_always_on": true, "supports_mid_conversation_system": true, @@ -57427,7 +57749,7 @@ "supports_output_config": true, "prompt_cache_min_tokens": 512, "supports_native_structured_output": true, - "source": "https://platform.claude.com/docs/en/models/mythos-5-1/overview" + "source": "https://platform.claude.com/docs/en/about-claude/pricing" }, "claude-mythos-preview": { "cache_creation_input_token_cost": 1.25e-05, @@ -57779,6 +58101,7 @@ }, "vertex_ai/gemini-3.5-live-translate-preview": { "input_cost_per_audio_token": 3.5e-06, + "input_cost_per_second": 8.83333333333e-05, "input_cost_per_token": 3.5e-06, "litellm_provider": "vertex_ai", "mode": "realtime", @@ -57881,14 +58204,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57897,14 +58223,17 @@ }, "fireworks_ai/accounts/fireworks/models/deepseek-v4p1-flash": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://fireworks.ai/models/deepseek-ai/deepseek-v4p1-flash", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -57920,26 +58249,29 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_tool_choice": true, "supports_vision": true }, "fireworks_ai/accounts/fireworks/models/kimi-k3": { "cache_read_input_token_cost": 3e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 3e-06, + "input_cost_per_token_priority": 3.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-05, + "output_cost_per_token_priority": 1.875e-05, "reasoning_effort_levels": [ "low", "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57948,14 +58280,17 @@ }, "fireworks_ai/deepseek-v4-flash-0731": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -57964,14 +58299,17 @@ }, "fireworks_ai/deepseek-v4p1-flash": { "cache_read_input_token_cost": 7e-09, + "cache_read_input_token_cost_priority": 8.75e-09, "input_cost_per_token": 2.2e-07, + "input_cost_per_token_priority": 2.75e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 393216, "max_tokens": 393216, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://fireworks.ai/models/deepseek-ai/deepseek-v4p1-flash", + "output_cost_per_token_priority": 8.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -57987,7 +58325,7 @@ "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 6.6e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_tool_choice": true, "supports_vision": true @@ -58026,19 +58364,22 @@ }, "fireworks_ai/kimi-k3": { "cache_read_input_token_cost": 3e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 3e-06, + "input_cost_per_token_priority": 3.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 131072, "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 1.5e-05, + "output_cost_per_token_priority": 1.875e-05, "reasoning_effort_levels": [ "low", "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58089,12 +58430,15 @@ }, "fireworks_ai/qwen3p8-max": { "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_priority": 3e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 9e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58110,7 +58454,7 @@ "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58142,7 +58486,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 2.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58158,7 +58502,7 @@ "max_tokens": 16384, "mode": "chat", "output_cost_per_token": 1.5e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58190,7 +58534,7 @@ "max_tokens": 32768, "mode": "chat", "output_cost_per_token": 2.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58199,12 +58543,15 @@ }, "fireworks_ai/accounts/fireworks/models/qwen3p8-max": { "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_priority": 3.75e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_priority": 3e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 262144, "mode": "chat", "output_cost_per_token": 6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 9e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58220,7 +58567,7 @@ "max_tokens": 131072, "mode": "chat", "output_cost_per_token": 6.6e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -58257,7 +58604,7 @@ "high", "max" ], - "source": "https://docs.fireworks.ai/serverless/pricing", + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -60955,14 +61302,17 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p3": { "cache_read_input_token_cost": 2.6e-07, + "cache_read_input_token_cost_priority": 3.25e-07, "input_cost_per_token": 1.4e-06, + "input_cost_per_token_priority": 1.75e-06, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 4.4e-06, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 5.5e-06, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_reasoning": true, "supports_response_schema": true, @@ -60971,13 +61321,16 @@ }, "fireworks_ai/accounts/fireworks/models/glm-5p3-flash": { "cache_read_input_token_cost": 3e-08, + "cache_read_input_token_cost_priority": 3.75e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_priority": 1.875e-07, "litellm_provider": "fireworks_ai", "max_input_tokens": 1048576, "max_tokens": 1048576, "mode": "chat", "output_cost_per_token": 5e-07, - "source": "https://docs.fireworks.ai/serverless/pricing", + "output_cost_per_token_priority": 6.25e-07, + "source": "https://api.fireworks.ai/v1/serverless/models", "supports_function_calling": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -61005,7 +61358,7 @@ "max_output_tokens": 40960, "max_tokens": 40960, "mode": "embedding", - "source": "https://docs.fireworks.ai/serverless/pricing" + "source": "https://api.fireworks.ai/v1/serverless/models" }, "zai/glm-5.2": { "cache_creation_input_token_cost": 0, @@ -61029,7 +61382,7 @@ "max_tokens": 1000000, "mode": "chat", "output_cost_per_token": 4.7e-07, - "source": "https://docs.together.ai/docs/serverless-models" + "source": "https://api.together.ai/v1/models" }, "together_ai/moonshotai/Kimi-K2.6": { "deprecation_date": "2026-08-19", @@ -61039,7 +61392,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/moonshotai/Kimi-K2.5-fp4": { "input_cost_per_token": 5e-07, @@ -61047,7 +61400,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/MiniMaxAI/MiniMax-M2.7": { "input_cost_per_token": 3e-07, @@ -61056,7 +61409,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 196608, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/zai-org/GLM-5": { "deprecation_date": "2026-06-22", @@ -61065,7 +61418,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 202752, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/zai-org/GLM-5.1": { "deprecation_date": "2026-07-10", @@ -61075,7 +61428,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 202752, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/deepseek-ai/DeepSeek-R1-0528": { "input_cost_per_token": 3e-06, @@ -61083,7 +61436,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 163840, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-Coder-Next-FP8": { "deprecation_date": "2026-05-14", @@ -61092,7 +61445,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-VL-32B-Instruct": { "deprecation_date": "2026-02-25", @@ -61101,7 +61454,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/Qwen3-VL-8B-Instruct": { "deprecation_date": "2026-04-16", @@ -61110,7 +61463,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Ministral-3-14B-Instruct-2512": { "input_cost_per_token": 2e-07, @@ -61118,7 +61471,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 262144, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/nvidia/NVIDIA-Nemotron-Nano-9B-v2": { "input_cost_per_token": 6e-08, @@ -61126,7 +61479,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 131072, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/mistralai/Mistral-7B-Instruct-v0.3": { "input_cost_per_token": 2e-07, @@ -61134,7 +61487,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 32768, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "together_ai/Qwen/QwQ-32B": { "deprecation_date": "2025-11-13", @@ -61143,7 +61496,7 @@ "litellm_provider": "together_ai", "max_input_tokens": 131072, "mode": "chat", - "source": "https://api.together.xyz/v1/models" + "source": "https://api.together.ai/v1/models" }, "cerebras/gemma-4-31b": { "input_cost_per_token": 9.9e-07, @@ -65270,5 +65623,260 @@ "supports_tool_choice": false, "supports_response_schema": true, "supports_vision": false + }, + "fireworks_ai/accounts/fireworks/routers/glm-5p3-fast": { + "cache_read_input_token_cost": 3.9e-07, + "input_cost_per_token": 2.1e-06, + "litellm_provider": "fireworks_ai", + "mode": "chat", + "output_cost_per_token": 6.6e-06, + "source": "https://api.fireworks.ai/v1/serverless/models" + }, + "together_ai/arcee-ai/trinity-mini": { + "input_cost_per_token": 4.5e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.5e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/deepseek-coder-33b-instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B": { + "input_cost_per_token": 2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B": { + "input_cost_per_token": 1.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B": { + "input_cost_per_token": 1.6e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.6e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/deepseek-ai/DeepSeek-V4.1-Flash": { + "cache_read_input_token_cost": 6e-09, + "input_cost_per_token": 3e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "vertex_ai/gemini-2.5-flash-native-audio": { + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "realtime", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-2.5-flash-preview-tts": { + "input_cost_per_token": 5e-07, + "input_cost_per_token_batches": 2.5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_audio_token": 1e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.1-flash-live-preview": { + "input_cost_per_audio_token": 3e-06, + "input_cost_per_second": 8.33333333333e-05, + "input_cost_per_token": 7.5e-07, + "litellm_provider": "vertex_ai", + "mode": "realtime", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 4.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.1-flash-tts-preview": { + "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_audio_token": 2e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.5-transcribe": { + "input_cost_per_audio_token": 2e-06, + "input_cost_per_second": 5e-05, + "litellm_provider": "vertex_ai", + "mode": "audio_transcription", + "output_cost_per_token": 1.2e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-3.5-transcribe-live": { + "input_cost_per_audio_token": 3.5e-06, + "input_cost_per_second": 8.33333333333e-05, + "litellm_provider": "vertex_ai", + "mode": "audio_transcription", + "output_cost_per_token": 2.1e-05, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-omni-1.1-flash": { + "input_cost_per_token": 1.5e-06, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 9e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemini-robotics-er-2": { + "cache_read_input_token_cost": 1e-07, + "input_cost_per_token": 1e-06, + "input_cost_per_token_batches": 5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 5e-06, + "output_cost_per_token_batches": 2.5e-06, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "vertex_ai/gemma-4-26b-a4b-it": { + "cache_read_input_token_cost": 1.5e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "vertex_ai", + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing" + }, + "together_ai/google/gemma-2-27b-it": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "gpt-5.5-cyber": { + "cache_read_input_token_cost": 1.25e-06, + "input_cost_per_token": 1.25e-05, + "litellm_provider": "openai", + "mode": "chat", + "output_cost_per_token": 7.5e-05, + "source": "https://developers.openai.com/api/docs/pricing" + }, + "gpt-rosalind-research": { + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 5e-06, + "litellm_provider": "openai", + "mode": "chat", + "output_cost_per_token": 2.5e-05, + "source": "https://developers.openai.com/api/docs/pricing" + }, + "together_ai/meta-llama/Llama-3-8b-chat-hf": { + "input_cost_per_token": 2e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.1-405B-Instruct": { + "input_cost_per_token": 3.5e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 3.5e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.2-1B-Instruct": { + "input_cost_per_token": 6e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Llama-3.2-3B-Instruct": { + "input_cost_per_token": 6e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Meta-Llama-3-70B-Instruct-Turbo": { + "input_cost_per_token": 8.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/meta-llama/Meta-Llama-3-8B-Instruct": { + "input_cost_per_token": 2e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO": { + "input_cost_per_token": 6e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF": { + "input_cost_per_token": 8.8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8.8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-1.5B-Instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 2e-08, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-72B-Instruct": { + "input_cost_per_token": 9e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 9e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2-VL-72B-Instruct": { + "input_cost_per_token": 1.2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-14B-Instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-72B-Instruct": { + "input_cost_per_token": 1.2e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-Coder-32B-Instruct": { + "input_cost_per_token": 8e-07, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-07, + "source": "https://api.together.ai/v1/models" + }, + "together_ai/Qwen/Qwen2.5-VL-72B-Instruct": { + "input_cost_per_token": 1.95e-06, + "litellm_provider": "together_ai", + "mode": "chat", + "output_cost_per_token": 8e-06, + "source": "https://api.together.ai/v1/models" } } From 76b26e41abfe6e72dd846e7272dcb45069b98b73 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 09:07:53 +0000 Subject: [PATCH 009/119] fix(router): cool down team deployments on 429 when a sibling serves the same public model Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/router.py | 10 ++++ litellm/router_utils/cooldown_handlers.py | 5 +- .../router_utils/test_cooldown_handlers.py | 50 +++++++++++++++++++ 3 files changed, 63 insertions(+), 2 deletions(-) diff --git a/litellm/router.py b/litellm/router.py index 8865543badd..4929b17f7fc 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -1553,6 +1553,16 @@ class Router: return False return sum(len(self.model_name_to_deployment_indices.get(member) or ()) for member in group.models) > 1 + def team_model_has_alternatives(self, deployment_id: str) -> bool: + deployment: Final = self.get_deployment(model_id=deployment_id) + if deployment is None: + return False + team_id: Final = deployment.model_info.team_id + public_model_name: Final = deployment.model_info.team_public_model_name + if team_id is None or public_model_name is None: + return False + return len(self.team_model_to_deployment_indices.get((team_id, public_model_name)) or ()) > 1 + _OVERRIDABLE_ROUTING_STRATEGIES: frozenset[str] = frozenset({"simple-shuffle", *_DEFAULT_SELECTOR_ATTR_BY_STRATEGY}) def _get_request_routing_strategy_override(self, request_kwargs: dict | None) -> str | None: diff --git a/litellm/router_utils/cooldown_handlers.py b/litellm/router_utils/cooldown_handlers.py index f722b6fd20c..027f0a9ca05 100644 --- a/litellm/router_utils/cooldown_handlers.py +++ b/litellm/router_utils/cooldown_handlers.py @@ -343,8 +343,9 @@ def _should_cooldown_deployment( model_group: Final = litellm_router_instance.get_model_group(id=deployment) is_single_deployment_model_group = False if model_group is not None and len(model_group) == 1: - is_single_deployment_model_group = not litellm_router_instance.routing_group_has_alternatives( - requested_model_group + is_single_deployment_model_group = not ( + litellm_router_instance.routing_group_has_alternatives(requested_model_group) + or litellm_router_instance.team_model_has_alternatives(deployment) ) ## CHECK DEPLOYMENT-LEVEL POLICY FIRST (overrides router-level) diff --git a/tests/test_litellm/router_utils/test_cooldown_handlers.py b/tests/test_litellm/router_utils/test_cooldown_handlers.py index 7ee0ed3701b..6f66bb863cc 100644 --- a/tests/test_litellm/router_utils/test_cooldown_handlers.py +++ b/tests/test_litellm/router_utils/test_cooldown_handlers.py @@ -437,3 +437,53 @@ class TestRoutingGroupCooldownAlternatives: ) is False ) + + +class TestTeamModelCooldownAlternatives: + def _router(self, team_deployments: int): + from litellm import Router + + return Router( + model_list=[ + { + "model_name": f"model_name_team-1_{i}", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, + "model_info": { + "id": f"team-deploy-{i}", + "team_id": "team-1", + "team_public_model_name": "team-gpt-4o-mini", + }, + } + for i in range(team_deployments) + ] + ) + + def test_429_on_team_deployment_with_sibling_cools_down(self): + from litellm.router_utils.cooldown_handlers import _should_cooldown_deployment + + router = self._router(team_deployments=2) + assert ( + _should_cooldown_deployment( + litellm_router_instance=router, + deployment="team-deploy-0", + exception_status=429, + original_exception=Exception("rate limited"), + requested_model_group="team-gpt-4o-mini", + ) + is True + ) + + def test_429_on_only_team_deployment_keeps_single_deployment_exemption(self): + from litellm.router_utils.cooldown_handlers import _should_cooldown_deployment + + router = self._router(team_deployments=1) + assert ( + _should_cooldown_deployment( + litellm_router_instance=router, + deployment="team-deploy-0", + exception_status=429, + original_exception=Exception("rate limited"), + requested_model_group="team-gpt-4o-mini", + ) + is False + ) From d0a846c8be5ec1ae9036254eeb04575f6b406921 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 09:27:33 +0000 Subject: [PATCH 010/119] test(router): cover team_model_has_alternatives directly in the mapped router test Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/test_router.py | 34 +++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+) diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index b8a0d70f5bc..7af60e01f7f 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -896,6 +896,40 @@ def test_arouter_test_team_model(): assert result is not None +def test_team_model_has_alternatives(): + def team_deployment(deployment_id: str, team_id: str, public_model_name: str): + return { + "model_name": f"model_name_{team_id}_{deployment_id}", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, + "model_info": { + "id": deployment_id, + "team_id": team_id, + "team_public_model_name": public_model_name, + }, + } + + router = litellm.Router( + model_list=[ + team_deployment("team-a-1", "team-a", "shared-model"), + team_deployment("team-a-2", "team-a", "shared-model"), + team_deployment("team-a-solo", "team-a", "solo-model"), + team_deployment("team-b-1", "team-b", "shared-model"), + { + "model_name": "plain-model", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, + "model_info": {"id": "plain-1"}, + }, + ], + ) + + assert router.team_model_has_alternatives("team-a-1") is True + assert router.team_model_has_alternatives("team-a-2") is True + assert router.team_model_has_alternatives("team-a-solo") is False + assert router.team_model_has_alternatives("team-b-1") is False + assert router.team_model_has_alternatives("plain-1") is False + assert router.team_model_has_alternatives("missing-deployment") is False + + def test_arouter_ignore_invalid_deployments(): """ Test that router.ignore_invalid_deployments is set to True From 9080f0904ad6bee6c5f10debdccdd1b064e1efe3 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 09:35:54 +0000 Subject: [PATCH 011/119] fix(router): ignore blocked siblings when checking team model cooldown alternatives Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/router.py | 4 +++- .../router_utils/test_cooldown_handlers.py | 18 +++++++++++++++++- tests/test_litellm/test_router.py | 6 +++++- 3 files changed, 25 insertions(+), 3 deletions(-) diff --git a/litellm/router.py b/litellm/router.py index 4929b17f7fc..63d5f86d46a 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -1561,7 +1561,9 @@ class Router: public_model_name: Final = deployment.model_info.team_public_model_name if team_id is None or public_model_name is None: return False - return len(self.team_model_to_deployment_indices.get((team_id, public_model_name)) or ()) > 1 + sibling_indices: Final = self.team_model_to_deployment_indices.get((team_id, public_model_name)) or () + routable_siblings: Final = self._filter_blocked_deployments([self.model_list[idx] for idx in sibling_indices]) + return len(routable_siblings) > 1 _OVERRIDABLE_ROUTING_STRATEGIES: frozenset[str] = frozenset({"simple-shuffle", *_DEFAULT_SELECTOR_ATTR_BY_STRATEGY}) diff --git a/tests/test_litellm/router_utils/test_cooldown_handlers.py b/tests/test_litellm/router_utils/test_cooldown_handlers.py index 6f66bb863cc..fdbfd618dab 100644 --- a/tests/test_litellm/router_utils/test_cooldown_handlers.py +++ b/tests/test_litellm/router_utils/test_cooldown_handlers.py @@ -440,7 +440,7 @@ class TestRoutingGroupCooldownAlternatives: class TestTeamModelCooldownAlternatives: - def _router(self, team_deployments: int): + def _router(self, team_deployments: int, blocked_ids: frozenset[str] = frozenset()): from litellm import Router return Router( @@ -452,6 +452,7 @@ class TestTeamModelCooldownAlternatives: "id": f"team-deploy-{i}", "team_id": "team-1", "team_public_model_name": "team-gpt-4o-mini", + "blocked": f"team-deploy-{i}" in blocked_ids, }, } for i in range(team_deployments) @@ -487,3 +488,18 @@ class TestTeamModelCooldownAlternatives: ) is False ) + + def test_429_with_only_a_blocked_sibling_keeps_single_deployment_exemption(self): + from litellm.router_utils.cooldown_handlers import _should_cooldown_deployment + + router = self._router(team_deployments=2, blocked_ids=frozenset({"team-deploy-1"})) + assert ( + _should_cooldown_deployment( + litellm_router_instance=router, + deployment="team-deploy-0", + exception_status=429, + original_exception=Exception("rate limited"), + requested_model_group="team-gpt-4o-mini", + ) + is False + ) diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 7af60e01f7f..7786bfb4788 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -897,7 +897,7 @@ def test_arouter_test_team_model(): def test_team_model_has_alternatives(): - def team_deployment(deployment_id: str, team_id: str, public_model_name: str): + def team_deployment(deployment_id: str, team_id: str, public_model_name: str, blocked: bool = False): return { "model_name": f"model_name_{team_id}_{deployment_id}", "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, @@ -905,6 +905,7 @@ def test_team_model_has_alternatives(): "id": deployment_id, "team_id": team_id, "team_public_model_name": public_model_name, + "blocked": blocked, }, } @@ -914,6 +915,8 @@ def test_team_model_has_alternatives(): team_deployment("team-a-2", "team-a", "shared-model"), team_deployment("team-a-solo", "team-a", "solo-model"), team_deployment("team-b-1", "team-b", "shared-model"), + team_deployment("team-c-1", "team-c", "paused-sibling-model"), + team_deployment("team-c-paused", "team-c", "paused-sibling-model", blocked=True), { "model_name": "plain-model", "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, @@ -926,6 +929,7 @@ def test_team_model_has_alternatives(): assert router.team_model_has_alternatives("team-a-2") is True assert router.team_model_has_alternatives("team-a-solo") is False assert router.team_model_has_alternatives("team-b-1") is False + assert router.team_model_has_alternatives("team-c-1") is False assert router.team_model_has_alternatives("plain-1") is False assert router.team_model_has_alternatives("missing-deployment") is False From e41b3bd13fa6415de7b4076dda82f673fac8b957 Mon Sep 17 00:00:00 2001 From: yassin Date: Sun, 13 Sep 2026 09:50:12 +0000 Subject: [PATCH 012/119] test(router): annotate return types of team cooldown test helpers Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/test_litellm/router_utils/test_cooldown_handlers.py | 6 ++---- tests/test_litellm/test_router.py | 4 +++- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/tests/test_litellm/router_utils/test_cooldown_handlers.py b/tests/test_litellm/router_utils/test_cooldown_handlers.py index fdbfd618dab..6fed4be5909 100644 --- a/tests/test_litellm/router_utils/test_cooldown_handlers.py +++ b/tests/test_litellm/router_utils/test_cooldown_handlers.py @@ -440,10 +440,8 @@ class TestRoutingGroupCooldownAlternatives: class TestTeamModelCooldownAlternatives: - def _router(self, team_deployments: int, blocked_ids: frozenset[str] = frozenset()): - from litellm import Router - - return Router( + def _router(self, team_deployments: int, blocked_ids: frozenset[str] = frozenset()) -> litellm.Router: + return litellm.Router( model_list=[ { "model_name": f"model_name_team-1_{i}", diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 7786bfb4788..c6b3b7fb8d6 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -897,7 +897,9 @@ def test_arouter_test_team_model(): def test_team_model_has_alternatives(): - def team_deployment(deployment_id: str, team_id: str, public_model_name: str, blocked: bool = False): + def team_deployment( + deployment_id: str, team_id: str, public_model_name: str, blocked: bool = False + ) -> DeploymentTypedDict: return { "model_name": f"model_name_{team_id}_{deployment_id}", "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test"}, From 0201ca60e7397912cf46e0a1bddf1d6befb86e0a Mon Sep 17 00:00:00 2001 From: ryan Date: Sun, 13 Sep 2026 23:09:27 +0000 Subject: [PATCH 013/119] fix(ui): move tags typed into key metadata JSON into the Tags field Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../templates/KeyEditViewControls.tsx | 41 +++++++++++++++++- .../templates/keyEditFieldNormalizers.test.ts | 42 +++++++++++++++++++ .../templates/keyEditFieldNormalizers.ts | 33 +++++++++++++++ .../key_edit_view.integration.test.tsx | 26 ++++++++++++ .../components/templates/key_edit_view.tsx | 22 ++++++---- 5 files changed, 156 insertions(+), 8 deletions(-) create mode 100644 ui/litellm-dashboard/src/components/templates/keyEditFieldNormalizers.test.ts diff --git a/ui/litellm-dashboard/src/components/templates/KeyEditViewControls.tsx b/ui/litellm-dashboard/src/components/templates/KeyEditViewControls.tsx index f5d5562ccfe..39542882798 100644 --- a/ui/litellm-dashboard/src/components/templates/KeyEditViewControls.tsx +++ b/ui/litellm-dashboard/src/components/templates/KeyEditViewControls.tsx @@ -1,12 +1,15 @@ import React from "react"; -import { Control } from "react-hook-form"; +import { Control, UseFormReturn } from "react-hook-form"; import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select"; +import { Textarea } from "@/components/ui/textarea"; import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"; import { CircleHelp } from "lucide-react"; import { FormField } from "@/components/shared/form/FormField"; +import { toast } from "@/lib/toast"; import AgentSelector from "../agent_management/AgentSelector"; import NumericalInput from "../shared/numerical_input"; import SkillSelector from "../skills/SkillSelector"; +import { moveTagsOutOfMetadataJson } from "./keyEditFieldNormalizers"; import { AgentsAndGroups, KeyEditFormValues } from "./keyEditFormValues"; export const labelWithHint = (label: React.ReactNode, hint: string): React.ReactNode => ( @@ -85,6 +88,42 @@ export const KeyAgentAndSkillFields = ({ ); +type KeyEditForm = Pick< + UseFormReturn, + "control" | "getValues" | "setValue" +>; + +export const moveMetadataTagsToTagsField = (form: KeyEditForm): void => { + const moved = moveTagsOutOfMetadataJson(form.getValues("metadata"), form.getValues("tags")); + if (moved === null) return; + form.setValue("metadata", moved.metadata, { shouldDirty: true }); + form.setValue("tags", moved.tags, { shouldDirty: true }); + if (moved.movedTags.length > 0) { + toast.info(`Moved ${moved.movedTags.join(", ")} from metadata to the Tags field`); + } +}; + +export const KeyMetadataField = ({ form }: { form: KeyEditForm }) => ( + + {(field) => ( +