diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index fc29a6e3a08..a1adc3071b5 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -4869,8 +4869,9 @@ class ProxyBaseLLMRequestProcessing: if stream_usage is None: return None service_tier: Final = obj.get("service_tier") + served_model: Final = litellm_logging_obj.model if litellm_logging_obj is not None else None cost_val: Final = ProxyBaseLLMRequestProcessing._streamed_usage_cost( - ModelResponse(usage=stream_usage), + ModelResponse(model=served_model, usage=stream_usage), model_name, service_tier if isinstance(service_tier, str) else None, litellm_logging_obj, diff --git a/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py b/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py index ae95c27950c..8127ffe558b 100644 --- a/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py +++ b/tests/unit/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py @@ -41,12 +41,13 @@ def _make_streaming_response(chunks): return mock -def _unarmed_logging_obj(): +def _unarmed_logging_obj(model: str = "unknown"): """Real Logging objects only carry _on_deferred_stream_complete when the proxy arms deferred dispatch; a bare MagicMock's auto-attribute is truthy and would spuriously trigger the deferral branch.""" obj = MagicMock() obj._on_deferred_stream_complete = None + obj.model = model return obj @@ -409,7 +410,7 @@ async def _collect_openai_passthrough_chunks(chunks, endpoint_type): async for chunk in PassThroughStreamingHandler.chunk_processor( response=response, request_body={"model": "gpt-4o-mini", "stream": True}, - litellm_logging_obj=_unarmed_logging_obj(), + litellm_logging_obj=_unarmed_logging_obj(model="gpt-4o-mini"), endpoint_type=endpoint_type, start_time=datetime(2026, 1, 1), passthrough_success_handler_obj=MagicMock(), diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index fdcbcbfcd05..b604841feb6 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -68,6 +68,7 @@ from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitErro from litellm.proxy.common_utils.sse_keepalive import ANTHROPIC_PING_SSE_CHUNK from litellm.proxy.dd_span_tagger import DDSpanTagger from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.llms.anthropic.chat.transformation import AnthropicConfig from litellm.proxy._types import Litellm_EntityType, ProxyErrorTypes, ProxyException from litellm.proxy._types import UserAPIKeyAuth as ProxyUserAPIKeyAuth from litellm.proxy.auth.fallback_budget import RouterFallbackBudgetCheck @@ -10148,6 +10149,8 @@ class TestInjectCostIntoUsageDict: discount streamed a ``usage.cost`` that disagreed with the callback's ``response_cost``.""" class _StubLoggingObj: + model = "claude-haiku-4-5" + def __init__(self, cost): self._cost = cost self.captured_result = None @@ -10184,6 +10187,8 @@ class TestInjectCostIntoUsageDict: def test_message_delta_falls_back_to_model_pricing_when_the_logging_obj_returns_no_cost(self): class _StubLoggingObj: + model = "claude-haiku-4-5" + def response_cost_calculator(self, result): return None @@ -10209,6 +10214,8 @@ class TestInjectCostIntoUsageDict: model-name pricing rather than propagating into the response body.""" class _StubLoggingObj: + model = "claude-haiku-4-5" + def response_cost_calculator(self, result): raise ValueError("no pricing for this deployment") @@ -10294,6 +10301,8 @@ class TestInjectCostIntoUsageDict: streaming /v1/chat/completions gets its negotiated price instead of sticker.""" class _StubLoggingObj: + model = "gpt-4o-mini" + def __init__(self, cost): self._cost = cost self.captured_result = None @@ -10322,6 +10331,8 @@ class TestInjectCostIntoUsageDict: def test_openai_chunk_falls_back_to_model_pricing_when_the_logging_obj_returns_no_cost(self): class _StubLoggingObj: + model = "gpt-4o-mini" + def response_cost_calculator(self, result): return None @@ -10337,6 +10348,60 @@ class TestInjectCostIntoUsageDict: assert result is not None assert result["usage"]["cost"] == pytest.approx(self._expected_cost("gpt-4o-mini", 11, 4)) + @pytest.mark.parametrize( + ("deployment_model", "served_model", "custom_llm_provider"), + [ + ("vertex_ai/claude-sonnet-5-5", "claude-sonnet-5-5", "vertex_ai"), + ("anthropic/claude-sonnet-5-5", "claude-sonnet-5-5", "anthropic"), + ("azure_ai/claude-sonnet-5-5", "claude-sonnet-5-5", "azure_ai"), + ], + ) + def test_message_delta_bills_search_like_the_logged_response_on_a_token_priced_claude_deployment( + self, local_model_cost_map: None, deployment_model: str, served_model: str, custom_llm_provider: str + ) -> None: + token_rates: Final = {"input_cost_per_token": 3e-6, "output_cost_per_token": 1.5e-5} + router: Final = Router( + model_list=[ + { + "model_name": "priced-claude", + "litellm_params": { + "model": deployment_model, + "api_key": "k", + "vertex_project": "p", + "vertex_location": "global", + }, + "model_info": token_rates, + } + ] + ) + deployment_id: Final = router.model_list[0]["model_info"]["id"] + logging_obj: Final = LiteLLMLoggingObj( + model=deployment_model, + messages=[{"role": "user", "content": "test"}], + stream=True, + call_type="anthropic_messages", + start_time=None, + litellm_call_id="token-priced-claude-search", + function_id="token-priced-claude-search", + ) + logging_obj.update_environment_variables( + litellm_params={"metadata": {"model_info": {"id": deployment_id, **token_rates}}}, optional_params={} + ) + logging_obj.model_call_details["custom_llm_provider"] = custom_llm_provider + wire_usage: Final = {"input_tokens": 1840, "output_tokens": 412, "server_tool_use": {"web_search_requests": 2}} + event: Final = {"type": "message_delta", "delta": {"stop_reason": "end_turn"}, "usage": wire_usage} + + result: Final = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, "priced-claude", logging_obj) + + stream_usage: Final = AnthropicConfig().calculate_usage(usage_object=wire_usage, reasoning_content=None) + logged_cost: Final = logging_obj.response_cost_calculator( + result=litellm.ModelResponse(model=served_model, usage=stream_usage) + ) + token_cost: Final = 1840 * token_rates["input_cost_per_token"] + 412 * token_rates["output_cost_per_token"] + assert logged_cost is not None and logged_cost > token_cost + assert result is not None + assert result["usage"]["cost"] == pytest.approx(logged_cost) + class TestProcessChunkWithCostInjection: @pytest.mark.parametrize("model", ["claude-opus-5-5", "vertex_ai/claude-opus-5-5"]) @@ -10403,6 +10468,8 @@ class TestProcessChunkWithCostInjection: monkeypatch.setattr(litellm, "include_cost_in_streaming_usage", True) class _StubLoggingObj: + model = "claude-haiku-4-5" + def response_cost_calculator(self, result): return 0.00042