fix(openai): bridge gpt-5.6+ tools to /v1/responses without reasoning_effort

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/openai/chat/gpt_5_transformation.py

@@ -102,18 +102,33 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
         return model_name.startswith("gpt-5.4")
 
     @classmethod
-    def is_model_gpt_5_4_plus_model(cls, model: str) -> bool:
-        """Check if the model is gpt-5.4 or newer (5.4, 5.5, 5.6, etc., including pro)."""
+    def _gpt_5_minor_version_at_least(cls, model: str, minimum: int) -> bool:
+        """Return True when a versioned gpt-5.<minor> name is at or above ``minimum``.
+
+        Non-versioned names (e.g. ``gpt-5``, ``gpt-5-codex``) and unparseable names
+        return False. Named or pro variants such as ``gpt-5.6-sol`` and ``gpt-5.4-pro``
+        are handled by taking the minor version before the first ``-``.
+        """
         model_name = model.split("/")[-1]
         if not model_name.startswith("gpt-5."):
             return False
         try:
             version_str = model_name.replace("gpt-5.", "").split("-")[0]
             major = version_str.split(".")[0]
-            return int(major) >= 4
+            return int(major) >= minimum
         except (ValueError, IndexError):
             return False
 
+    @classmethod
+    def is_model_gpt_5_4_plus_model(cls, model: str) -> bool:
+        """Check if the model is gpt-5.4 or newer (5.4, 5.5, 5.6, etc., including pro)."""
+        return cls._gpt_5_minor_version_at_least(model, 4)
+
+    @classmethod
+    def is_model_gpt_5_6_plus_model(cls, model: str) -> bool:
+        """Check if the model is gpt-5.6 or newer (5.6, 5.7, etc., including named/pro variants)."""
+        return cls._gpt_5_minor_version_at_least(model, 6)
+
     @classmethod
     def _supports_reasoning_effort_level(cls, model: str, level: str) -> bool:
         """Check if the model supports a specific reasoning_effort level.

litellm/main.py

@@ -1022,7 +1022,11 @@ def responses_api_bridge_check(
     # ``reasoningSummary`` in ``extra_body``) must be bridged; Chat Completions rejects
     # those keys.
     #
-    # - gpt-5.4+: tools + reasoning_effort (original) or any reasoning-summary alias.
+    # - gpt-5.6+: tools alone bridge, even without reasoning_effort. OpenAI applies a
+    #   default reasoning_effort server-side for this family, so tools on Chat
+    #   Completions are rejected outright (see issue #33221).
+    # - gpt-5.4/5.5: tools + reasoning_effort, or any reasoning-summary alias. Tools
+    #   alone (no reasoning_effort) stay on Chat Completions.
     # - Older GPT-5 names (e.g. ``gpt-5``, ``gpt-5.1``): bridge only when a reasoning
     #   summary alias is present with ``reasoning_effort`` (tools alone stay on chat).
     if (
@@ -1030,8 +1034,13 @@ def responses_api_bridge_check(
         and model_info.get("mode") != "responses"
         and OpenAIGPT5Config.is_model_gpt_5_model(model)
         and not OpenAIGPT5Config.is_model_gpt_5_search_model(model)
-        and reasoning_effort is not None
-        and (reasoning_summary is not None or (OpenAIGPT5Config.is_model_gpt_5_4_plus_model(model) and tools))
+        and (
+            (
+                reasoning_effort is not None
+                and (reasoning_summary is not None or (OpenAIGPT5Config.is_model_gpt_5_4_plus_model(model) and tools))
+            )
+            or (OpenAIGPT5Config.is_model_gpt_5_6_plus_model(model) and tools)
+        )
     ):
         model_info["mode"] = "responses"
         model = model.replace("responses/", "")

tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py

@@ -155,6 +155,44 @@ class TestOpenAIGPT5ConfigIsModelGpt54PlusModel:
         ), f"Expected '{model}' NOT to be classified as gpt-5.4-or-newer"
 
 
+# Models that are gpt-5.6 or newer. main.py bridges tools-only (no reasoning_effort)
+# requests to /v1/responses for exactly this set, so gpt-5.4/5.5 must land on the
+# False side while the gpt-5.6 family (including named variants) lands on True.
+GPT5_6_PLUS_MODELS = [
+    "gpt-5.6",
+    "gpt-5.6-sol",
+    "gpt-5.6-terra",
+    "gpt-5.6-luna",
+    "openai/gpt-5.6-sol",
+]
+
+GPT5_PRE_5_6_MODELS = [
+    "gpt-5",
+    "gpt-5.1",
+    "gpt-5.3",
+    "gpt-5.4",
+    "gpt-5.4-pro",
+    "gpt-5.5",
+    "gpt-5.5-pro",
+    "gpt-4o",
+]
+
+
+class TestOpenAIGPT5ConfigIsModelGpt56PlusModel:
+
+    @pytest.mark.parametrize("model", GPT5_6_PLUS_MODELS)
+    def test_gpt5_6_plus_models_are_classified_as_5_6_plus(self, model: str):
+        assert OpenAIGPT5Config.is_model_gpt_5_6_plus_model(
+            model
+        ), f"Expected '{model}' to be classified as gpt-5.6-or-newer"
+
+    @pytest.mark.parametrize("model", GPT5_PRE_5_6_MODELS)
+    def test_pre_5_6_models_are_not_classified_as_5_6_plus(self, model: str):
+        assert not OpenAIGPT5Config.is_model_gpt_5_6_plus_model(
+            model
+        ), f"Expected '{model}' NOT to be classified as gpt-5.6-or-newer"
+
+
 # ---------------------------------------------------------------------------
 # AzureOpenAIGPT5Config
 # ---------------------------------------------------------------------------

tests/test_litellm/test_main.py

@@ -898,6 +898,74 @@ def test_responses_api_bridge_check_gpt_5_tools_without_summary_stays_chat():
     assert model_info.get("mode") != "responses"
 
 
+@pytest.mark.parametrize(
+    "model",
+    ["gpt-5.6", "gpt-5.6-sol", "gpt-5.6-luna", "gpt-5.6-terra"],
+)
+@pytest.mark.parametrize("custom_llm_provider", ["openai", "azure"])
+def test_responses_api_bridge_check_gpt_5_6_tools_without_reasoning_routes_to_responses(
+    model, custom_llm_provider
+):
+    """gpt-5.6+ with tools but no reasoning_effort must bridge to Responses API.
+
+    Regression test for https://github.com/BerriAI/litellm/issues/33221
+    OpenAI applies a default reasoning_effort server-side for the gpt-5.6 family,
+    so function tools on /v1/chat/completions are rejected with a 400.
+    """
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, returned_model = responses_api_bridge_check(
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort=None,
+        )
+
+    assert returned_model == model
+    assert model_info.get("mode") == "responses"
+
+
+def test_responses_api_bridge_check_gpt_5_6_tools_plus_reasoning_routes_to_responses():
+    """gpt-5.6 with tools and explicit reasoning_effort still bridges (explicit-effort path)."""
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, model = responses_api_bridge_check(
+            model="gpt-5.6-sol",
+            custom_llm_provider="openai",
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort="high",
+        )
+
+    assert model == "gpt-5.6-sol"
+    assert model_info.get("mode") == "responses"
+
+
+@pytest.mark.parametrize("model", ["gpt-5.4", "gpt-5.5", "gpt-5.5-pro"])
+def test_responses_api_bridge_check_gpt_5_4_and_5_5_tools_without_reasoning_stay_chat(model):
+    """gpt-5.4/5.5 tools-only (no reasoning_effort) must NOT bridge; only 5.6+ does.
+
+    Guards against widening the gpt-5.6 tools-only bridge back onto older families,
+    which is the behavior the maintainer regression tests above pin down.
+    """
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, returned_model = responses_api_bridge_check(
+            model=model,
+            custom_llm_provider="openai",
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort=None,
+        )
+
+    assert returned_model == model
+    assert model_info.get("mode") != "responses"
+
+
 @patch("litellm.completion_extras.responses_api_bridge.completion")
 def test_gpt_5_4_responses_bridge_preserves_reasoning_summary_dict(
     mock_responses_completion,