From e0027bdc53c2612f6a8a7264d7d74ff145f81731 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 17:12:30 +0300
Subject: [PATCH] fix(responses): reject unsupported reasoning effort

---
 .../llms/openai/responses/transformation.py   | 32 ++++++++++++---
 .../test_reasoning_effort_capability.py       | 40 +++++++++++++++++++
 2 files changed, 67 insertions(+), 5 deletions(-)
 create mode 100644 tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index eac844a790d..b4299d2c907 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -121,9 +121,8 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     ) -> dict:
         """No mapping applied since inputs are in OpenAI spec already.
 
-        GPT-5 models have restrictions on temperature (only temperature=1
-        is accepted unless reasoning_effort='none' on models that support it).
-        Apply the same validation used by the chat completions path.
+        GPT-5 models have restrictions on temperature and reasoning effort.
+        Apply the same capability checks used by the chat completions path.
         """
         params: Final = dict(response_api_optional_params)
 
@@ -131,10 +130,33 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             params["max_output_tokens"] = self._enforce_min_max_output_tokens(params.get("max_output_tokens"))
 
         if self._is_gpt_5_model(model=model):
+            reasoning: Final = params.get("reasoning") or {}
+            effort: Final = reasoning.get("effort") if isinstance(reasoning, dict) else None
+            if isinstance(effort, str):
+                from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config
+
+                unsupported_effort: Final = (
+                    effort == "xhigh" and not OpenAIGPT5Config._supports_reasoning_effort_level(model, effort)
+                ) or (
+                    effort in ("minimal", "low")
+                    and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(model, effort)
+                )
+                if unsupported_effort:
+                    if drop_params or litellm.drop_params:
+                        updated_reasoning: Final = dict(reasoning)
+                        updated_reasoning.pop("effort", None)
+                        if updated_reasoning:
+                            params["reasoning"] = updated_reasoning
+                        else:
+                            params.pop("reasoning", None)
+                    else:
+                        raise litellm.UnsupportedParamsError(
+                            message=f"reasoning.effort={effort} is not supported for this model.",
+                            status_code=400,
+                        )
+
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
-                reasoning: Final = params.get("reasoning") or {}
-                effort: Final = reasoning.get("effort") if isinstance(reasoning, dict) else None
                 supports_none: Final = self._supports_reasoning_effort_none(model=model)
                 if supports_none and self._effort_resolves_to_none(model, effort):
                     pass  # flexible temperature allowed
diff --git a/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py b/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
new file mode 100644
index 00000000000..e055f413faf
--- /dev/null
+++ b/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
@@ -0,0 +1,40 @@
+import pytest
+
+import litellm
+from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
+
+
+@pytest.mark.parametrize("effort", ["minimal", "low"])
+def test_rejects_explicitly_unsupported_lower_reasoning_effort(effort: str) -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    with pytest.raises(litellm.UnsupportedParamsError, match=f"reasoning.effort={effort}"):
+        config.map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": effort}},
+            model="gpt-5.5-pro",
+            drop_params=False,
+        )
+
+
+def test_keeps_supported_reasoning_effort() -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "medium"}},
+        model="gpt-5.5-pro",
+        drop_params=False,
+    )
+
+    assert result["reasoning"] == {"effort": "medium"}
+
+
+def test_drop_params_removes_only_unsupported_effort() -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "minimal", "summary": "detailed"}},
+        model="gpt-5.5-pro",
+        drop_params=True,
+    )
+
+    assert result["reasoning"] == {"summary": "detailed"}