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 01/19] 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"}

From 7a295933a5fb2157b11cede760c6ded7a27aa6db Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 17:25:49 +0300
Subject: [PATCH 02/19] fix(responses): validate temperature after effort drop

---
 litellm/llms/openai/responses/transformation.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index b4299d2c907..fcb6760d3b3 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -157,8 +157,12 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
+                transformed_reasoning: Final = params.get("reasoning") or {}
+                transformed_effort: Final = (
+                    transformed_reasoning.get("effort") if isinstance(transformed_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):
+                if supports_none and self._effort_resolves_to_none(model, transformed_effort):
                     pass  # flexible temperature allowed
                 elif drop_params or litellm.drop_params:
                     params.pop("temperature", None)

From f985141e7cdfda45188124a7cc8cc1932d215d75 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 17:26:11 +0300
Subject: [PATCH 03/19] test(responses): cover effort drop temperature
 interaction

---
 .../test_reasoning_effort_capability.py        | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

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
index e055f413faf..2eef324eacd 100644
--- a/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
+++ b/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
@@ -1,6 +1,7 @@
 import pytest
 
 import litellm
+from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 
 
@@ -38,3 +39,20 @@ def test_drop_params_removes_only_unsupported_effort() -> None:
     )
 
     assert result["reasoning"] == {"summary": "detailed"}
+
+
+def test_drop_params_revalidates_temperature_after_effort_drop(monkeypatch: pytest.MonkeyPatch) -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    monkeypatch.setattr(OpenAIGPT5Config, "_supports_reasoning_effort_level", lambda model, level: False)
+    monkeypatch.setattr(config, "_supports_reasoning_effort_none", lambda model: True)
+    monkeypatch.setattr(config, "_effort_resolves_to_none", lambda model, effort: effort is None)
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "xhigh"}, "temperature": 0.5},
+        model="gpt-5-test",
+        drop_params=True,
+    )
+
+    assert "reasoning" not in result
+    assert result["temperature"] == 0.5

From 111bc6dfb1a542dbb28b8e0b5bfa0effce3c9623 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 17:46:51 +0300
Subject: [PATCH 04/19] chore(responses): satisfy type discipline gate

---
 litellm/llms/openai/responses/transformation.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index fcb6760d3b3..f0932e3dd11 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -143,7 +143,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 )
                 if unsupported_effort:
                     if drop_params or litellm.drop_params:
-                        updated_reasoning: Final = dict(reasoning)
+                        updated_reasoning: Final = dict(reasoning)  # mutable-ok: local request payload copy before removing unsupported effort
                         updated_reasoning.pop("effort", None)
                         if updated_reasoning:
                             params["reasoning"] = updated_reasoning
@@ -157,7 +157,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
-                transformed_reasoning: Final = params.get("reasoning") or {}
+                transformed_reasoning: Final = params.get("reasoning") or {}  # mutable-ok: local request view for transformed effort validation
                 transformed_effort: Final = (
                     transformed_reasoning.get("effort") if isinstance(transformed_reasoning, dict) else None
                 )
@@ -365,7 +365,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                         else item_data.get("reasoning_content", "")
                     )
 
-                # Create ResponseReasoningItem object from the item data
+                # Create ResponsesReasoningItem object with item data
                 reasoning_item: Final = ResponseReasoningItem(**item_data)
 
                 # Convert back to dict with exclude_none=True to exclude None fields

From a1c0bdb97a99090134438fe56990aa9d58b9741f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 17:48:49 +0300
Subject: [PATCH 05/19] chore(responses): restore existing comment

---
 litellm/llms/openai/responses/transformation.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index f0932e3dd11..47bbcbeda4e 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -365,7 +365,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                         else item_data.get("reasoning_content", "")
                     )
 
-                # Create ResponsesReasoningItem object with item data
+                # Create ResponseReasoningItem object from the item data
                 reasoning_item: Final = ResponseReasoningItem(**item_data)
 
                 # Convert back to dict with exclude_none=True to exclude None fields

From dbdb157bee5b8f959abced45a4d2c2186867153b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 18:08:50 +0300
Subject: [PATCH 06/19] chore(responses): keep lint comments formatted

---
 litellm/llms/openai/responses/transformation.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 47bbcbeda4e..7301f12fe21 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -143,7 +143,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 )
                 if unsupported_effort:
                     if drop_params or litellm.drop_params:
-                        updated_reasoning: Final = dict(reasoning)  # mutable-ok: local request payload copy before removing unsupported effort
+                        updated_reasoning: Final = dict(reasoning)  # mutable-ok: local copy
                         updated_reasoning.pop("effort", None)
                         if updated_reasoning:
                             params["reasoning"] = updated_reasoning
@@ -157,7 +157,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
             temperature: Final = params.get("temperature")
             if temperature is not None and temperature != 1:
-                transformed_reasoning: Final = params.get("reasoning") or {}  # mutable-ok: local request view for transformed effort validation
+                transformed_reasoning: Final = params.get("reasoning") or {}  # mutable-ok: local view
                 transformed_effort: Final = (
                     transformed_reasoning.get("effort") if isinstance(transformed_reasoning, dict) else None
                 )
@@ -794,4 +794,4 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         response._hidden_params["additional_headers"] = processed_headers
         response._hidden_params["headers"] = raw_response_headers
 
-        return response
+        return response
\ No newline at end of file

From 88a444f8f599e661f48985e58aae72a28019f188 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 18:10:57 +0300
Subject: [PATCH 07/19] chore(responses): restore final newline

---
 litellm/llms/openai/responses/transformation.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 7301f12fe21..2b71038c14a 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -794,4 +794,4 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         response._hidden_params["additional_headers"] = processed_headers
         response._hidden_params["headers"] = raw_response_headers
 
-        return response
\ No newline at end of file
+        return response

From 031cad4cfd9838ae997382d682c136d9e1cd872c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 18:40:50 +0300
Subject: [PATCH 08/19] chore(responses): satisfy basedpyright gate

---
 litellm/llms/openai/responses/transformation.py | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 2b71038c14a..61aa35c1f36 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -136,10 +136,15 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config
 
                 unsupported_effort: Final = (
-                    effort == "xhigh" and not OpenAIGPT5Config._supports_reasoning_effort_level(model, effort)
+                    effort == "xhigh"
+                    and not OpenAIGPT5Config._supports_reasoning_effort_level(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
+                        model, effort
+                    )
                 ) or (
                     effort in ("minimal", "low")
-                    and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(model, effort)
+                    and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
+                        model, effort
+                    )
                 )
                 if unsupported_effort:
                     if drop_params or litellm.drop_params:

From 97e201b1c604b294bf221c3bf4486f28aad35931 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 18:49:38 +0300
Subject: [PATCH 09/19] chore(responses): suppress helper call diagnostics

---
 litellm/llms/openai/responses/transformation.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 61aa35c1f36..05d05c9e135 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -138,12 +138,12 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 unsupported_effort: Final = (
                     effort == "xhigh"
                     and not OpenAIGPT5Config._supports_reasoning_effort_level(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model, effort
+                        model, effort  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
                     )
                 ) or (
                     effort in ("minimal", "low")
                     and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model, effort
+                        model, effort  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
                     )
                 )
                 if unsupported_effort:

From d6514effaba9b456edbe21efc89206e6f2cfe59e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 19:15:33 +0300
Subject: [PATCH 10/19] chore(responses): format helper suppressions

---
 litellm/llms/openai/responses/transformation.py | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 05d05c9e135..74b0a2fff6d 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -138,12 +138,14 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 unsupported_effort: Final = (
                     effort == "xhigh"
                     and not OpenAIGPT5Config._supports_reasoning_effort_level(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model, effort  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
+                        model,
+                        effort,  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
                     )
                 ) or (
                     effort in ("minimal", "low")
                     and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model, effort  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
+                        model,
+                        effort,  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
                     )
                 )
                 if unsupported_effort:

From 3a30409bec55796537e7bd78a7de776f6bab5c9d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 19:30:01 +0300
Subject: [PATCH 11/19] fix(responses): keep reasoning effort typed

---
 .../llms/openai/responses/transformation.py   | 64 ++++++++++---------
 1 file changed, 34 insertions(+), 30 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 74b0a2fff6d..a8caf332ca0 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -81,6 +81,25 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
         return OpenAIGPT5Config.effort_resolves_to_none(model, effort)
 
+    @staticmethod
+    def _is_unsupported_reasoning_effort(model: str, effort: str | None) -> bool:
+        """Apply the GPT-5 reasoning-effort capability flags used by chat completions."""
+        from litellm.utils import _is_explicitly_disabled_factory, _supports_factory
+
+        if effort == "xhigh":
+            return not _supports_factory(
+                model=model,
+                custom_llm_provider=None,
+                key="supports_xhigh_reasoning_effort",
+            )
+        if effort in ("minimal", "low"):
+            return _is_explicitly_disabled_factory(
+                model=model,
+                custom_llm_provider=None,
+                key=f"supports_{effort}_reasoning_effort",
+            )
+        return False
+
     @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.
@@ -130,46 +149,31 @@ 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
+            reasoning: Final = response_api_optional_params.get("reasoning")
+            effort: Final = reasoning.get("effort") if reasoning is not None else None
+            unsupported_effort: Final = self._is_unsupported_reasoning_effort(model, effort)
+            should_drop_effort: Final = unsupported_effort and (drop_params or litellm.drop_params)
 
-                unsupported_effort: Final = (
-                    effort == "xhigh"
-                    and not OpenAIGPT5Config._supports_reasoning_effort_level(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model,
-                        effort,  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
-                    )
-                ) or (
-                    effort in ("minimal", "low")
-                    and OpenAIGPT5Config._is_reasoning_effort_level_explicitly_disabled(  # pyright: ignore[reportPrivateUsage, reportArgumentType, reportCallIssue]  # shared GPT-5 capability gate
-                        model,
-                        effort,  # pyright: ignore[reportArgumentType, reportCallIssue]  # shared helper call
-                    )
-                )
-                if unsupported_effort:
-                    if drop_params or litellm.drop_params:
-                        updated_reasoning: Final = dict(reasoning)  # mutable-ok: local copy
+            if unsupported_effort:
+                if should_drop_effort:
+                    if reasoning is not None:
+                        updated_reasoning: Final = reasoning.copy()  # mutable-ok: local copy
                         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,
-                        )
+                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:
-                transformed_reasoning: Final = params.get("reasoning") or {}  # mutable-ok: local view
-                transformed_effort: Final = (
-                    transformed_reasoning.get("effort") if isinstance(transformed_reasoning, dict) else None
-                )
+                effective_effort: Final = None if should_drop_effort else effort
                 supports_none: Final = self._supports_reasoning_effort_none(model=model)
-                if supports_none and self._effort_resolves_to_none(model, transformed_effort):
+                if supports_none and self._effort_resolves_to_none(model, effective_effort):
                     pass  # flexible temperature allowed
                 elif drop_params or litellm.drop_params:
                     params.pop("temperature", None)

From 7dc5ae4d27c6ee7911c61aa9cfc55f49268d50f5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 20:01:33 +0300
Subject: [PATCH 12/19] fix(responses): use public reasoning capability helper

---
 .../llms/openai/responses/transformation.py    | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index a8caf332ca0..54ce6a366e6 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -84,20 +84,16 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     @staticmethod
     def _is_unsupported_reasoning_effort(model: str, effort: str | None) -> bool:
         """Apply the GPT-5 reasoning-effort capability flags used by chat completions."""
-        from litellm.utils import _is_explicitly_disabled_factory, _supports_factory
+        from litellm.router_utils.reasoning_effort_capability import (
+            declared_reasoning_efforts_for_model,
+        )
 
+        lookup_model: Final = model.removeprefix("openai/")
+        declared: Final = declared_reasoning_efforts_for_model(lookup_model, "openai")
         if effort == "xhigh":
-            return not _supports_factory(
-                model=model,
-                custom_llm_provider=None,
-                key="supports_xhigh_reasoning_effort",
-            )
+            return declared is None or effort not in declared
         if effort in ("minimal", "low"):
-            return _is_explicitly_disabled_factory(
-                model=model,
-                custom_llm_provider=None,
-                key=f"supports_{effort}_reasoning_effort",
-            )
+            return declared is not None and effort not in declared
         return False
 
     @staticmethod

From 96a9fb6e1a758f0f59b4504a89bcb6422ad1da77 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Onat=20=C3=96zmen?= <onatozmen44@gmail.com>
Date: Sun, 30 Aug 2026 20:28:23 +0300
Subject: [PATCH 13/19] fix(responses): share GPT-5 effort capability gate

---
 litellm/llms/openai/chat/gpt_5_transformation.py |  9 +++++++++
 litellm/llms/openai/responses/transformation.py  | 12 ++----------
 2 files changed, 11 insertions(+), 10 deletions(-)

diff --git a/litellm/llms/openai/chat/gpt_5_transformation.py b/litellm/llms/openai/chat/gpt_5_transformation.py
index 0223be300b0..bfaf396c7fb 100644
--- a/litellm/llms/openai/chat/gpt_5_transformation.py
+++ b/litellm/llms/openai/chat/gpt_5_transformation.py
@@ -156,6 +156,15 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
             key=f"supports_{level}_reasoning_effort",
         )
 
+    @classmethod
+    def is_reasoning_effort_unsupported(cls, model: str, level: str | None) -> bool:
+        """Return whether GPT-5 capability metadata rejects this effort level."""
+        if level == "xhigh":
+            return not cls._supports_reasoning_effort_level(model, level)
+        if level in ("minimal", "low"):
+            return cls._is_reasoning_effort_level_explicitly_disabled(model, level)
+        return False
+
     @classmethod
     def effort_resolves_to_none(cls, model: str, effective_effort: str | None) -> bool:
         """Whether this request's reasoning effort ends up as "none", which is the single
diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 54ce6a366e6..8dd8b574ee8 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -84,17 +84,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     @staticmethod
     def _is_unsupported_reasoning_effort(model: str, effort: str | None) -> bool:
         """Apply the GPT-5 reasoning-effort capability flags used by chat completions."""
-        from litellm.router_utils.reasoning_effort_capability import (
-            declared_reasoning_efforts_for_model,
-        )
+        from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config
 
-        lookup_model: Final = model.removeprefix("openai/")
-        declared: Final = declared_reasoning_efforts_for_model(lookup_model, "openai")
-        if effort == "xhigh":
-            return declared is None or effort not in declared
-        if effort in ("minimal", "low"):
-            return declared is not None and effort not in declared
-        return False
+        return OpenAIGPT5Config.is_reasoning_effort_unsupported(model, effort)
 
     @staticmethod
     def _enforce_min_max_output_tokens(max_output_tokens: "int | None") -> "int | None":

From 0764a0a923bf1e7e3cd6689929baa03b4f6a50a2 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Mon, 14 Sep 2026 00:12:15 +0300
Subject: [PATCH 14/19] fix(responses): dispatch effort capability check by
 provider and drop test monkeypatching - Azure responses config reads Azure
 map entry (OpenAI entry disables minimal on gpt-5.4, Azure enables it);
 temperature revalidation test now uses real gpt-5.1 map data

---
 .../llms/azure/responses/transformation.py    |  9 +++++
 .../test_reasoning_effort_capability.py       | 40 +++++++++++++++----
 2 files changed, 42 insertions(+), 7 deletions(-)

diff --git a/litellm/llms/azure/responses/transformation.py b/litellm/llms/azure/responses/transformation.py
index 7fe12138ebc..696840b0d17 100644
--- a/litellm/llms/azure/responses/transformation.py
+++ b/litellm/llms/azure/responses/transformation.py
@@ -38,6 +38,15 @@ class AzureOpenAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
     def _effort_resolves_to_none(model: str, effort: str | None) -> bool:
         return AzureOpenAIGPT5Config.effort_resolves_to_none(model, effort)
 
+    @staticmethod
+    def _is_unsupported_reasoning_effort(model: str, effort: str | None) -> bool:
+        """Read the effort capability from the Azure map entry, like the sibling checks above.
+
+        The inherited OpenAI lookup would resolve a bare Azure deployment name (e.g. ``gpt-5.4``)
+        against the OpenAI entry, which can explicitly disable an effort Azure supports.
+        """
+        return AzureOpenAIGPT5Config.is_reasoning_effort_unsupported(model, effort)
+
     def get_supported_openai_params(self, model: str) -> list:
         """
         Azure Responses API does not support context_management (compaction).
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
index 2eef324eacd..cdb8f3ee4bc 100644
--- a/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
+++ b/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
@@ -1,7 +1,7 @@
 import pytest
 
 import litellm
-from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config
+from litellm.llms.azure.responses.transformation import AzureOpenAIResponsesAPIConfig
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 
 
@@ -41,18 +41,44 @@ def test_drop_params_removes_only_unsupported_effort() -> None:
     assert result["reasoning"] == {"summary": "detailed"}
 
 
-def test_drop_params_revalidates_temperature_after_effort_drop(monkeypatch: pytest.MonkeyPatch) -> None:
-    config = OpenAIResponsesAPIConfig()
+def test_drop_params_revalidates_temperature_after_effort_drop() -> None:
+    """Dropping an unsupported effort revalidates temperature against the emptied effort.
 
-    monkeypatch.setattr(OpenAIGPT5Config, "_supports_reasoning_effort_level", lambda model, level: False)
-    monkeypatch.setattr(config, "_supports_reasoning_effort_none", lambda model: True)
-    monkeypatch.setattr(config, "_effort_resolves_to_none", lambda model, effort: effort is None)
+    `gpt-5.1` really declares `xhigh` unsupported while supporting (and defaulting to)
+    `none`, so no capability stubbing is needed: the dropped effort resolves to `none`
+    and a valid non-default temperature survives.
+    """
+    config = OpenAIResponsesAPIConfig()
 
     result = config.map_openai_params(
         response_api_optional_params={"reasoning": {"effort": "xhigh"}, "temperature": 0.5},
-        model="gpt-5-test",
+        model="gpt-5.1",
         drop_params=True,
     )
 
     assert "reasoning" not in result
     assert result["temperature"] == 0.5
+
+
+def test_azure_config_uses_azure_capability_map_for_effort() -> None:
+    """A bare Azure deployment name must resolve against the Azure map entry.
+
+    OpenAI's `gpt-5.4` entry explicitly disables `minimal` while Azure's enables it;
+    the OpenAI lookup would wrongly reject an effort Azure supports.
+    """
+    openai_config = OpenAIResponsesAPIConfig()
+    with pytest.raises(litellm.UnsupportedParamsError, match="reasoning.effort=minimal"):
+        openai_config.map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": "minimal"}},
+            model="gpt-5.4",
+            drop_params=False,
+        )
+
+    azure_config = AzureOpenAIResponsesAPIConfig()
+    result = azure_config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "minimal"}},
+        model="gpt-5.4",
+        drop_params=False,
+    )
+
+    assert result["reasoning"] == {"effort": "minimal"}

From 14d81858dbae967553dd618499b922ed11cc91d2 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Mon, 14 Sep 2026 00:30:11 +0300
Subject: [PATCH 15/19] fix(tests): wrap pytest.raises match in re.escape to
 satisfy RUF043

---
 .../llms/openai/responses/test_reasoning_effort_capability.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

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
index cdb8f3ee4bc..585a0ea3b5f 100644
--- a/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
+++ b/tests/test_litellm/llms/openai/responses/test_reasoning_effort_capability.py
@@ -1,3 +1,5 @@
+import re
+
 import pytest
 
 import litellm
@@ -67,7 +69,7 @@ def test_azure_config_uses_azure_capability_map_for_effort() -> None:
     the OpenAI lookup would wrongly reject an effort Azure supports.
     """
     openai_config = OpenAIResponsesAPIConfig()
-    with pytest.raises(litellm.UnsupportedParamsError, match="reasoning.effort=minimal"):
+    with pytest.raises(litellm.UnsupportedParamsError, match=re.escape("reasoning.effort=minimal")):
         openai_config.map_openai_params(
             response_api_optional_params={"reasoning": {"effort": "minimal"}},
             model="gpt-5.4",

From d1cea94f497c9a3a3610e8c6176d9328be656c95 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Mon, 14 Sep 2026 01:21:38 +0300
Subject: [PATCH 16/19] fix(auth): cache prefetched org entries with the
 management TTL - the 5s default fuse expired before slow requests reached
 get_org_object, sending them to the DB (proxy-behavior auth prefetch test
 failed under load with dead_db TypeError)

---
 litellm/proxy/auth/auth_object_prefetch.py                | 8 +++++---
 .../test_litellm/proxy/auth/test_auth_object_prefetch.py  | 4 ++--
 2 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/litellm/proxy/auth/auth_object_prefetch.py b/litellm/proxy/auth/auth_object_prefetch.py
index 52e26e885c9..503223e7d87 100644
--- a/litellm/proxy/auth/auth_object_prefetch.py
+++ b/litellm/proxy/auth/auth_object_prefetch.py
@@ -14,7 +14,6 @@ from pydantic import BaseModel, TypeAdapter, ValidationError
 
 from litellm._logging import verbose_proxy_logger
 from litellm.caching.redis_cache import RedisCache
-from litellm.constants import DEFAULT_IN_MEMORY_TTL
 from litellm.models.organization import LiteLLM_OrganizationTable
 from litellm.models.team import LiteLLM_TeamTableCachedObj
 from litellm.models.team_membership import LiteLLM_TeamMembership
@@ -190,14 +189,17 @@ def _iter_entries(refs: AuthObjectRefs, management_ttl: float) -> Iterator[_Cach
             None,
         )
     if refs.organization_id is not None:
+        # Organization entries use the management TTL like every other prefetched object: a
+        # 5s fuse expires before slow requests reach the getters, sending them back to
+        # the DB the prefetch was meant to spare.
         yield _CacheEntry(
-            f"org_id:{refs.organization_id}", "organization_row", LiteLLM_OrganizationTable, DEFAULT_IN_MEMORY_TTL
+            f"org_id:{refs.organization_id}", "organization_row", LiteLLM_OrganizationTable, management_ttl
         )
         yield _CacheEntry(
             f"org_id:{refs.organization_id}:with_budget",
             "organization_row",
             LiteLLM_OrganizationTable,
-            DEFAULT_IN_MEMORY_TTL,
+            management_ttl,
         )
     if refs.project_id is not None:
         yield _CacheEntry(f"project_id:{refs.project_id}", "project_row", LiteLLM_ProjectTableCachedObj, management_ttl)
diff --git a/tests/test_litellm/proxy/auth/test_auth_object_prefetch.py b/tests/test_litellm/proxy/auth/test_auth_object_prefetch.py
index 0fd0dda3017..58761e270a0 100644
--- a/tests/test_litellm/proxy/auth/test_auth_object_prefetch.py
+++ b/tests/test_litellm/proxy/auth/test_auth_object_prefetch.py
@@ -181,8 +181,8 @@ async def test_cold_regime_is_one_mget_one_query_and_the_getters_never_touch_io_
     assert sets == sorted(
         [
             f"SET {TEAM_ID}_{USER_ID} ttl=5",
-            f"SET org_id:{ORG_ID} ttl=5",
-            f"SET org_id:{ORG_ID}:with_budget ttl=5",
+            f"SET org_id:{ORG_ID} ttl=60",
+            f"SET org_id:{ORG_ID}:with_budget ttl=60",
             f"SET {USER_ID} ttl=60",
             f"SET team_id:{TEAM_ID} ttl=60",
             f"SET team_membership:{USER_ID}:{TEAM_ID} ttl=None",

From 7d6550e1bacfceffcb2756f4d6504178545b5bc1 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Mon, 14 Sep 2026 01:27:23 +0300
Subject: [PATCH 17/19] chore(ui): sync stale schema.d.ts soft_budget removal
 (matches generator output; unblocks Check UI API Types Sync now that the
 branch touches litellm/proxy/)

---
 ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 --
 1 file changed, 2 deletions(-)

diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 7eadaa6c991..839aa52fa84 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -16781,7 +16781,6 @@ export interface paths {
          *     - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *     - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *     - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *     - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *     - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *     - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
@@ -16887,7 +16886,6 @@ export interface paths {
          *         - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *         - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *         - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *         - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *         - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *         - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *         - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)

From dce9f6f2760a58ff35b460a224f0817791070c42 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Fri, 25 Sep 2026 15:06:19 +0300
Subject: [PATCH 18/19] \fix(lint): narrow reasoning to dict before copy to
 satisfy basedpyright budget"

---
 litellm/llms/openai/responses/transformation.py | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index 26179aa1140..ec91be473aa 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -256,12 +256,13 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
 
             if unsupported_effort:
                 if should_drop_effort:
-                    updated_reasoning: Final = dict(reasoning)  # mutable-ok: local copy
-                    updated_reasoning.pop("effort", None)
-                    if updated_reasoning:
-                        params["reasoning"] = updated_reasoning
-                    else:
-                        params.pop("reasoning", None)
+                    if isinstance(reasoning, dict):
+                        updated_reasoning: Final = reasoning.copy()  # mutable-ok: local copy
+                        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.",

From b9ffdf819ef727c6b77d9e09142e612782753f20 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Sat, 26 Sep 2026 17:41:04 +0300
Subject: [PATCH 19/19] fix(responses): remove unused suppression

---
 litellm/llms/openai/responses/transformation.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index ec91be473aa..0d523507a2a 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -257,7 +257,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             if unsupported_effort:
                 if should_drop_effort:
                     if isinstance(reasoning, dict):
-                        updated_reasoning: Final = reasoning.copy()  # mutable-ok: local copy
+                        updated_reasoning: Final = reasoning.copy()
                         updated_reasoning.pop("effort", None)
                         if updated_reasoning:
                             params["reasoning"] = updated_reasoning