From ed366aafbe12c126765ee8ee2073cf751918ecb1 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 28 Jul 2026 16:12:25 +0000 Subject: [PATCH] fix(anthropic): preserve prompt cache tokens in messages to responses api usage Also map gpt-5.6 flex/priority cache_creation rates into ModelInfo so cache writes are not billed at the standard rate on those service tiers --- .../responses_adapters/streaming_iterator.py | 33 +++++---------- .../responses_adapters/transformation.py | 32 +++++++++------ litellm/types/utils.py | 4 ++ litellm/utils.py | 4 ++ .../llm_cost_calc/test_llm_cost_calc_utils.py | 41 +++++++++++++++++++ ...t_responses_adapters_streaming_iterator.py | 29 +++++++++++++ .../test_responses_adapters_transformation.py | 34 +++++++++++++-- ui/litellm-dashboard/src/lib/http/schema.d.ts | 8 ++++ 8 files changed, 147 insertions(+), 38 deletions(-) diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py index 4fd49a35417..c3f0c8912ca 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py @@ -7,6 +7,9 @@ from typing import Any, AsyncIterator, Dict from litellm import verbose_logger from litellm._uuid import uuid +from litellm.types.llms.anthropic_messages.anthropic_response import AnthropicUsage + +from .transformation import LiteLLMAnthropicToResponsesAPIAdapter class AnthropicResponsesStreamWrapper: @@ -226,24 +229,17 @@ class AnthropicResponsesStreamWrapper: event.get("response") if isinstance(event, dict) else None ) stop_reason = "end_turn" - input_tokens = 0 - output_tokens = 0 - cache_creation_tokens = 0 - cache_read_tokens = 0 + anthropic_usage: AnthropicUsage = AnthropicUsage(input_tokens=0, output_tokens=0) if response_obj is not None: status = getattr(response_obj, "status", None) if status == "incomplete": stop_reason = "max_tokens" - usage = getattr(response_obj, "usage", None) - if usage is not None: - input_tokens = getattr(usage, "input_tokens", 0) or 0 - output_tokens = getattr(usage, "output_tokens", 0) or 0 - cache_creation_tokens = getattr(usage, "input_tokens_details", None) # type: ignore[assignment] - cache_read_tokens = getattr(usage, "output_tokens_details", None) # type: ignore[assignment] - # Prefer direct cache fields if present - cache_creation_tokens = int(getattr(usage, "cache_creation_input_tokens", 0) or 0) - cache_read_tokens = int(getattr(usage, "cache_read_input_tokens", 0) or 0) + anthropic_usage = ( + LiteLLMAnthropicToResponsesAPIAdapter.translate_responses_api_usage_to_anthropic_usage( + getattr(response_obj, "usage", None) + ) + ) # Check if tool_use was in the output to override stop_reason if response_obj is not None: @@ -256,20 +252,11 @@ class AnthropicResponsesStreamWrapper: stop_reason = "tool_use" break - usage_delta: Dict[str, Any] = { - "input_tokens": input_tokens, - "output_tokens": output_tokens, - } - if cache_creation_tokens: - usage_delta["cache_creation_input_tokens"] = cache_creation_tokens - if cache_read_tokens: - usage_delta["cache_read_input_tokens"] = cache_read_tokens - self._chunk_queue.append( { "type": "message_delta", "delta": {"stop_reason": stop_reason, "stop_sequence": None}, - "usage": usage_delta, + "usage": dict(anthropic_usage), } ) self._chunk_queue.append({"type": "message_stop"}) diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py index 172e54de98e..8b75b0447fa 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py @@ -29,7 +29,7 @@ from litellm.types.llms.anthropic_messages.anthropic_response import ( AnthropicMessagesResponse, AnthropicUsage, ) -from litellm.types.llms.openai import ResponsesAPIResponse +from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse class LiteLLMAnthropicToResponsesAPIAdapter: @@ -38,6 +38,24 @@ class LiteLLMAnthropicToResponsesAPIAdapter: converts Responses API responses back to Anthropic format. """ + @staticmethod + def translate_responses_api_usage_to_anthropic_usage( + raw_usage: Optional[ResponseAPIUsage], + ) -> AnthropicUsage: + """Map Responses API usage onto Anthropic usage, where ``input_tokens`` + excludes the cache-read and cache-write tokens reported alongside it. + """ + if raw_usage is None: + return AnthropicUsage(input_tokens=0, output_tokens=0) + + from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import ( + LiteLLMAnthropicMessagesAdapter, + ) + from litellm.responses.utils import ResponseAPILoggingUtils + + chat_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(raw_usage) + return LiteLLMAnthropicMessagesAdapter._translate_openai_usage_to_anthropic_usage(chat_usage) + # ------------------------------------------------------------------ # # Request translation: Anthropic -> Responses API # # ------------------------------------------------------------------ # @@ -396,8 +414,6 @@ class LiteLLMAnthropicToResponsesAPIAdapter: ResponseReasoningItem, ) - from litellm.types.llms.openai import ResponseAPIUsage - content: List[Dict[str, Any]] = [] stop_reason: AnthropicFinishReason = "end_turn" @@ -463,15 +479,7 @@ class LiteLLMAnthropicToResponsesAPIAdapter: if response.status == "incomplete": stop_reason = "max_tokens" - # usage - raw_usage: Optional[ResponseAPIUsage] = response.usage - input_tokens = int(getattr(raw_usage, "input_tokens", 0) or 0) - output_tokens = int(getattr(raw_usage, "output_tokens", 0) or 0) - - anthropic_usage = AnthropicUsage( - input_tokens=input_tokens, - output_tokens=output_tokens, - ) + anthropic_usage = self.translate_responses_api_usage_to_anthropic_usage(response.usage) return AnthropicMessagesResponse( id=response.id, diff --git a/litellm/types/utils.py b/litellm/types/utils.py index e4dfac48141..04388edaf05 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -198,6 +198,8 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False): input_cost_per_token_flex: Optional[float] # OpenAI flex service tier pricing input_cost_per_token_priority: Optional[float] # OpenAI priority service tier pricing cache_creation_input_token_cost: Optional[float] + cache_creation_input_token_cost_flex: Optional[float] # OpenAI flex service tier pricing + cache_creation_input_token_cost_priority: Optional[float] # OpenAI priority service tier pricing cache_creation_input_token_cost_above_200k_tokens: Optional[float] cache_creation_input_token_cost_above_1hr: Optional[float] cache_read_input_token_cost: Optional[float] @@ -3087,6 +3089,8 @@ class CustomPricingLiteLLMParams(BaseModel): input_cost_per_token_flex: Optional[float] = None input_cost_per_token_priority: Optional[float] = None cache_creation_input_token_cost: Optional[float] = None + cache_creation_input_token_cost_flex: Optional[float] = None + cache_creation_input_token_cost_priority: Optional[float] = None cache_creation_input_token_cost_above_1hr: Optional[float] = None cache_creation_input_token_cost_above_200k_tokens: Optional[float] = None cache_creation_input_audio_token_cost: Optional[float] = None diff --git a/litellm/utils.py b/litellm/utils.py index 944bb61d5e7..3296b39e708 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -5407,6 +5407,10 @@ def _get_model_info_helper( input_cost_per_token_flex=_model_info.get("input_cost_per_token_flex", None), input_cost_per_token_priority=_model_info.get("input_cost_per_token_priority", None), cache_creation_input_token_cost=_model_info.get("cache_creation_input_token_cost", None), + cache_creation_input_token_cost_flex=_model_info.get("cache_creation_input_token_cost_flex", None), + cache_creation_input_token_cost_priority=_model_info.get( + "cache_creation_input_token_cost_priority", None + ), cache_creation_input_token_cost_above_200k_tokens=_model_info.get( "cache_creation_input_token_cost_above_200k_tokens", None ), diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py index d282e656ce8..d1b488190a0 100644 --- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py +++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py @@ -2399,3 +2399,44 @@ def test_generic_cost_per_token_gemini_35_flash_lite(): ) assert prompt_cost == pytest.approx(0.0003) assert completion_cost == pytest.approx(0.00125) + + +@pytest.mark.parametrize( + "service_tier,input_rate,cache_read_rate,cache_write_rate,output_rate", + [ + ("flex", 2.5e-6, 2.5e-7, 3.125e-6, 1.5e-5), + ("priority", 1e-5, 1e-6, 1.25e-5, 6e-5), + ], +) +def test_service_tier_cache_creation_rates_for_gpt_5_6( + _local_model_cost_map, + service_tier, + input_rate, + cache_read_rate, + cache_write_rate, + output_rate, +): + """Regression: gpt-5.6 publishes cache_creation_input_token_cost_flex/_priority, so a + flex or priority request must bill cache writes at that tier's rate instead of falling + back to the standard 6.25e-6 rate.""" + usage = Usage( + prompt_tokens=10_000, + completion_tokens=500, + total_tokens=10_500, + prompt_tokens_details=PromptTokensDetailsWrapper( + cached_tokens=6_000, + cache_write_tokens=3_000, + text_tokens=1_000, + ), + ) + + prompt_cost, completion_cost = generic_cost_per_token( + model="gpt-5.6-sol", + usage=usage, + custom_llm_provider="openai", + service_tier=service_tier, + ) + + expected_prompt = 1_000 * input_rate + 6_000 * cache_read_rate + 3_000 * cache_write_rate + assert prompt_cost == pytest.approx(expected_prompt, rel=1e-9) + assert completion_cost == pytest.approx(500 * output_rate, rel=1e-9) diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py index 9b5197d9028..73b58e71009 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py @@ -6,6 +6,7 @@ Tests for AnthropicResponsesStreamWrapper import asyncio import os import sys +from types import SimpleNamespace sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../../.."))) @@ -130,3 +131,31 @@ class TestProcessEventTextDeltaWithoutOutputItemAdded: ("content_block_start", 0), ("content_block_delta", 0), ] + + +class TestResponseCompletedUsage: + """The Anthropic ``message_delta`` usage must report cache reads/writes and + exclude them from ``input_tokens``, so spend is not billed at the uncached + input rate.""" + + def test_response_completed_usage_carries_cache_tokens(self): + from litellm.types.llms.openai import ResponseAPIUsage + + response = SimpleNamespace( + status="completed", + output=[], + usage=ResponseAPIUsage( + input_tokens=4017, + input_tokens_details={"cached_tokens": 4004, "cache_write_tokens": 10}, + output_tokens=5, + total_tokens=4022, + ), + ) + chunks = _process_all([{"type": "response.completed", "response": response}]) + message_delta = next(c for c in chunks if c["type"] == "message_delta") + assert message_delta["usage"] == { + "input_tokens": 3, + "output_tokens": 5, + "cache_creation_input_tokens": 10, + "cache_read_input_tokens": 4004, + } diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py index 606ff39b35e..77b6f902368 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py @@ -20,6 +20,7 @@ from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transfo LiteLLMAnthropicToResponsesAPIAdapter, ) from litellm.types.llms.anthropic import AnthropicMessagesRequest +from litellm.types.llms.openai import ResponseAPIUsage def _make_request(**overrides) -> AnthropicMessagesRequest: @@ -823,11 +824,19 @@ def _make_mock_response( model: str = "gpt-4o", input_tokens: int = 100, output_tokens: int = 50, + cached_tokens: int = 0, + cache_write_tokens: int = 0, ) -> MagicMock: """Build a minimal mock ResponsesAPIResponse.""" - usage = MagicMock() - usage.input_tokens = input_tokens - usage.output_tokens = output_tokens + usage = ResponseAPIUsage( + input_tokens=input_tokens, + input_tokens_details={ + "cached_tokens": cached_tokens, + "cache_write_tokens": cache_write_tokens, + }, + output_tokens=output_tokens, + total_tokens=input_tokens + output_tokens, + ) resp = MagicMock() resp.id = response_id @@ -961,6 +970,25 @@ class TestTranslateResponse: assert result["usage"]["input_tokens"] == 200 assert result["usage"]["output_tokens"] == 75 + def test_cache_tokens_mapped_to_anthropic_usage(self): + """Cache reads/writes reported by the Responses API must survive the + Anthropic mapping, and input_tokens must exclude them so spend is not + billed at the uncached input rate.""" + response = _make_mock_response( + output=[_make_output_message(["OK"])], + input_tokens=4017, + output_tokens=5, + cached_tokens=4004, + cache_write_tokens=10, + ) + result: Any = _ADAPTER.translate_response(response) + assert result["usage"] == { + "input_tokens": 3, + "output_tokens": 5, + "cache_creation_input_tokens": 10, + "cache_read_input_tokens": 4004, + } + def test_model_and_id_preserved(self): """Model and response ID from the Responses API are forwarded.""" response = _make_mock_response( diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 541d7a17ae1..b9559113b84 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -25822,6 +25822,10 @@ export interface components { cache_creation_input_token_cost_above_1hr?: number | null; /** Cache Creation Input Token Cost Above 200K Tokens */ cache_creation_input_token_cost_above_200k_tokens?: number | null; + /** Cache Creation Input Token Cost Flex */ + cache_creation_input_token_cost_flex?: number | null; + /** Cache Creation Input Token Cost Priority */ + cache_creation_input_token_cost_priority?: number | null; /** Cache Read Input Audio Token Cost */ cache_read_input_audio_token_cost?: number | null; /** Cache Read Input Token Cost */ @@ -33912,6 +33916,10 @@ export interface components { cache_creation_input_token_cost_above_1hr?: number | null; /** Cache Creation Input Token Cost Above 200K Tokens */ cache_creation_input_token_cost_above_200k_tokens?: number | null; + /** Cache Creation Input Token Cost Flex */ + cache_creation_input_token_cost_flex?: number | null; + /** Cache Creation Input Token Cost Priority */ + cache_creation_input_token_cost_priority?: number | null; /** Cache Read Input Audio Token Cost */ cache_read_input_audio_token_cost?: number | null; /** Cache Read Input Token Cost */