From 4ef5db7c91ccfb4b690d811baf7cfad4129ab7ae Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 16:36:05 -0700
Subject: [PATCH] fix(responses): drop unsupported reasoning param for openai
 non-reasoning models

---
 .../llms/openai/responses/transformation.py   | 29 ++++++++++++
 .../test_openai_responses_transformation.py   | 46 +++++++++++++++++++
 2 files changed, 75 insertions(+)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 2fa44cfc2e3..99ce158c4e2 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -75,6 +75,19 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
         return OpenAIGPT5Config.effort_resolves_to_none(model, effort)
 
+    @staticmethod
+    def _is_o_series_name(model: str) -> bool:
+        base: Final = model.split("/")[-1]
+        return len(base) > 1 and base[0] == "o" and base[1].isdigit()
+
+    def _supports_reasoning_param(self, model: str) -> bool:
+        if self._is_gpt_5_model(model=model) or self._is_o_series_name(model=model):
+            return True
+        base: Final = model.split("/")[-1]
+        if base not in litellm.open_ai_chat_completion_models:
+            return True
+        return litellm.supports_reasoning(model=base, custom_llm_provider=self.custom_llm_provider.value)
+
     @staticmethod
     def _enforce_min_max_output_tokens(max_output_tokens: "int | None") -> "int | None":
         """Raise sub-minimum max_output_tokens up to the OpenAI Responses API minimum.
@@ -124,6 +137,22 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         if "max_output_tokens" in params:
             params["max_output_tokens"] = self._enforce_min_max_output_tokens(params.get("max_output_tokens"))
 
+        if (
+            self.custom_llm_provider == LlmProviders.OPENAI
+            and params.get("reasoning") is not None
+            and not self._supports_reasoning_param(model=model)
+        ):
+            if drop_params or litellm.drop_params:
+                params.pop("reasoning", None)
+            else:
+                raise litellm.UnsupportedParamsError(
+                    message=(
+                        f"{model} doesn't support the `reasoning` parameter. "
+                        "To drop unsupported params set `litellm.drop_params = True`"
+                    ),
+                    status_code=400,
+                )
+
         if self._is_gpt_5_model(model=model):
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
index e314b94444b..66d22cf8fb0 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
@@ -1626,3 +1626,49 @@ class TestResponsesSurfaceSharesTheEffortRule:
             drop_params=True,
         )
         assert ("temperature" in mapped) is temperature_survives
+
+
+class TestReasoningFollowsModelSupport:
+    """Responses API clients like Codex send `reasoning` on every request, and OpenAI 400s it
+    on non-reasoning models like gpt-4o. drop_params must strip it there, the same way the
+    chat completions surface already strips reasoning_effort for those models.
+    """
+
+    @pytest.mark.parametrize(
+        "model, reasoning_survives",
+        [
+            ("gpt-4o", False),
+            ("gpt-4.1", False),
+            ("gpt-4o-mini", False),
+            ("gpt-5.6", True),
+            ("o3", True),
+            ("o3-deep-research", True),
+            ("codex-mini-latest", True),
+            ("computer-use-preview", True),
+        ],
+    )
+    def test_drop_params_strips_reasoning_by_model(self, local_model_cost_map, model, reasoning_survives):
+        mapped = OpenAIResponsesAPIConfig().map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": "medium", "summary": "auto"}},
+            model=model,
+            drop_params=True,
+        )
+        assert ("reasoning" in mapped) is reasoning_survives
+
+    def test_without_drop_params_the_error_is_litellms_400(self, local_model_cost_map, monkeypatch):
+        monkeypatch.setattr(litellm, "drop_params", False)
+        with pytest.raises(litellm.UnsupportedParamsError) as excinfo:
+            OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": "medium"}},
+                model="gpt-4o",
+                drop_params=False,
+            )
+        assert excinfo.value.status_code == 400
+
+    def test_azure_deployments_keep_reasoning(self, local_model_cost_map):
+        mapped = AzureOpenAIResponsesAPIConfig().map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": "medium"}},
+            model="my-o3-deployment",
+            drop_params=True,
+        )
+        assert mapped["reasoning"] == {"effort": "medium"}