From 3831e66d2bcd4e458ba4a5dd6cfa5095636e4c7f Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 13 Aug 2026 01:32:13 +0000 Subject: [PATCH] fix(budget_reservation): don't reserve budget on token counting routes Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../spend_tracking/budget_reservation.py | 12 ++++- .../proxy/test_budget_reservation.py | 47 +++++++++++++++++++ 2 files changed, 58 insertions(+), 1 deletion(-) diff --git a/litellm/proxy/spend_tracking/budget_reservation.py b/litellm/proxy/spend_tracking/budget_reservation.py index 58a85171cc7..7b62fd44d09 100644 --- a/litellm/proxy/spend_tracking/budget_reservation.py +++ b/litellm/proxy/spend_tracking/budget_reservation.py @@ -144,6 +144,16 @@ async def _apply_over_budget_reservation_policy( ) +_UNBILLED_ROUTES: Final[frozenset[str]] = frozenset({"/models", "/v1/models", "/utils/token_counter"}) +_UNBILLED_ROUTE_SUFFIXES: Final[tuple[str, ...]] = ("/v1/messages/count_tokens", ":countTokens") + + +def _is_unbilled_route(route: str) -> bool: + """Routes that never emit a cost-tracking callback. Reserving budget for them + is a permanent leak: nothing ever reconciles or releases the reservation.""" + return route in _UNBILLED_ROUTES or route.endswith(_UNBILLED_ROUTE_SUFFIXES) + + async def reserve_budget_for_request( request_body: dict, route: str, @@ -161,7 +171,7 @@ async def reserve_budget_for_request( ) -> dict | None: if valid_token is None or not RouteChecks.is_llm_api_route(route=route): return None - if route in {"/models", "/v1/models", "/utils/token_counter"}: + if _is_unbilled_route(route): return None if get_model_from_request(request_body, route, llm_router=llm_router) is None: return None diff --git a/tests/test_litellm/proxy/test_budget_reservation.py b/tests/test_litellm/proxy/test_budget_reservation.py index 34adb4d2091..7bdc73abf32 100644 --- a/tests/test_litellm/proxy/test_budget_reservation.py +++ b/tests/test_litellm/proxy/test_budget_reservation.py @@ -2583,3 +2583,50 @@ async def test_streaming_slow_path_processes_and_yields_chunk(spend_counter_stat assert received == [{"content": "hi"}] streaming_logging_obj.async_post_call_streaming_hook.assert_awaited_once() + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "route", + [ + "/v1/messages/count_tokens", + "/anthropic/v1/messages/count_tokens", + "/v1beta/models/gemini-2.5-pro:countTokens", + "/models/gemini-2.5-pro:countTokens", + ], +) +async def test_token_counting_routes_never_reserve_budget(spend_counter_state, route): + """Token counting is free and never fires a cost callback, so a reservation + there is never reconciled and permanently bricks the key's spend counter.""" + counter_cache, key_cache = spend_counter_state + proxy_logging_obj = ProxyLogging(user_api_key_cache=key_cache) + valid_token = UserAPIKeyAuth( + token="key-count-tokens", + spend=0.0, + max_budget=0.01, + ) + + with patch( + "litellm.proxy.spend_tracking.budget_reservation.estimate_request_max_cost", + return_value=0.01, + ): + for _ in range(2): + assert ( + await reserve_budget_for_request( + request_body=_request_body(), + route=route, + llm_router=None, + valid_token=valid_token, + team_object=None, + user_object=None, + prisma_client=None, + user_api_key_cache=key_cache, + proxy_logging_obj=proxy_logging_obj, + ) + is None + ) + + assert counter_cache.in_memory_cache.get_cache(key="spend:key:key-count-tokens") is None + + # a real completion on the same key is still budget enforced + assert await _reserve(valid_token, 0.01, key_cache, proxy_logging_obj) is not None