diff --git a/litellm/llms/anthropic/pass_through/adapters/handler.py b/litellm/llms/anthropic/pass_through/adapters/handler.py index 01e781485a0..977b84fa3ea 100644 --- a/litellm/llms/anthropic/pass_through/adapters/handler.py +++ b/litellm/llms/anthropic/pass_through/adapters/handler.py @@ -401,6 +401,7 @@ class LiteLLMMessagesToCompletionTransformationHandler: completion_kwargs: _CompletionKwargs, *, thinking: Mapping[str, object] | None, + use_chat_completions_url: bool = False, ) -> None: """ When users call `litellm.anthropic.messages.*` with a non-Anthropic model and @@ -423,7 +424,7 @@ class LiteLLMMessagesToCompletionTransformationHandler: except Exception: custom_llm_provider = None - if custom_llm_provider != "openai": + if custom_llm_provider != "openai" or use_chat_completions_url: return if not isinstance(thinking, dict) or thinking.get("type") != "enabled": @@ -705,6 +706,7 @@ class LiteLLMMessagesToCompletionTransformationHandler: LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed( completion_kwargs, thinking=thinking, + use_chat_completions_url=litellm.use_chat_completions_url_for_anthropic_messages, ) chat_reasoning: Final = LiteLLMMessagesToCompletionTransformationHandler._chat_target_reasoning( diff --git a/tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py b/tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py index f4920505722..8dca368384a 100644 --- a/tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py +++ b/tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py @@ -427,6 +427,25 @@ class TestThinkingParameterTransformation: class TestThinkingSummaryPreservation: """Tests for thinking.summary preservation and reasoning_auto_summary flag.""" + @pytest.mark.parametrize("use_chat_completions_url", [True, False]) + def test_thinking_respects_chat_completions_url_override( + self, monkeypatch: pytest.MonkeyPatch, use_chat_completions_url: bool + ) -> None: + from litellm.llms.anthropic.pass_through.adapters.handler import ( + LiteLLMMessagesToCompletionTransformationHandler, + ) + + monkeypatch.setattr(litellm, "use_chat_completions_url_for_anthropic_messages", use_chat_completions_url) + completion_kwargs, _ = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs( + model="openai/gpt-5.6", + max_tokens=1024, + messages=[{"role": "user", "content": "Reply with OK"}], + thinking={"type": "enabled", "budget_tokens": 5000}, + ) + + expected_model: Final = "openai/gpt-5.6" if use_chat_completions_url else "openai/responses/gpt-5.6" + assert completion_kwargs["model"] == expected_model + def test_thinking_summary_concise_preserved_for_openai(self): """User-provided summary='concise' should not be replaced with 'detailed'.""" from litellm.llms.anthropic.pass_through.adapters.handler import (