fix(anthropic): keep sending reasoning_effort none on the chat route to an openai/ deployment the cost map does not list

The router registers an unlisted deployment as a stub row with no mode, which made its reasoning levels read as known-empty, so thinking disabled stopped sending none on the chat-completions route. The levels now read as unknown for such a row. The Responses bridge keeps sending no reasoning there, since the Responses transformation answers 400 to reasoning.effort on a row without supports_reasoning.

litellm/llms/anthropic/messages/responses_adapters/handler.py

@@ -52,6 +52,7 @@ from ..convertible_blocks import reject_malformed_tool_result_text
 from ..safeguards import with_safeguard_results
 from ..utils import (
     REASONING_EFFORT_OFF,
+    cost_map_row_lacks_reasoning,
     litellm_logging_obj_from_kwargs,
     local_model_name,
     maps_bare_output_effort,
@@ -276,7 +277,9 @@ def _reasoning_for_target(
         return reasoning
     if effort == REASONING_EFFORT_OFF:
         off_level: Final = reasoning_off_level(model, custom_llm_provider, drop_params=drop_params)
-        return None if off_level is None else {**reasoning_params, "effort": off_level}
+        if off_level is None or cost_map_row_lacks_reasoning(model, custom_llm_provider):
+            return None
+        return {**reasoning_params, "effort": off_level}
     if drop_params and effort not in _KNOWN_REASONING_EFFORTS:
         verbose_logger.warning(
             "Dropping reasoning.effort=%r for %s, no reasoning API defines it, because drop_params is set",

litellm/llms/anthropic/messages/utils.py

@@ -92,6 +92,7 @@ _OBJECT_LIST: Final = TypeAdapter(list[object])
 
 
 _NO_SETTINGS: Final[Mapping[str, object]] = MappingProxyType({})
+_NO_OPERATOR_MODEL_INFO: Final[Mapping[str, object]] = MappingProxyType({})
 
 
 def proxy_general_settings() -> Mapping[str, object]:
@@ -242,12 +243,17 @@ def _mapped_model_info(model: str, custom_llm_provider: str | None) -> ModelInfo
 
 
 def model_reasoning_efforts(model: str, custom_llm_provider: str | None) -> tuple[str, ...] | None:
-    from litellm.litellm_core_utils.reasoning_effort_capability import resolve_supported_reasoning_efforts
+    from litellm.litellm_core_utils.reasoning_effort_capability import (
+        deployment_is_catalog_mapped,
+        resolve_supported_reasoning_efforts,
+    )
 
     model_info: Final = _mapped_model_info(model, custom_llm_provider)
     if model_info is None:
         return None
-    return resolve_supported_reasoning_efforts(model_info, deployment_is_mapped=True)
+    return resolve_supported_reasoning_efforts(
+        model_info, deployment_is_mapped=deployment_is_catalog_mapped(model_info, _NO_OPERATOR_MODEL_INFO)
+    )
 
 
 def _lowest_tier_instead_of_off(levels: tuple[str, ...] | None) -> str | None:
@@ -256,6 +262,11 @@ def _lowest_tier_instead_of_off(levels: tuple[str, ...] | None) -> str | None:
     return next((tier for tier in _LOWEST_REASONING_TIERS if tier in levels), None)
 
 
+def cost_map_row_lacks_reasoning(model: str, custom_llm_provider: str | None) -> bool:
+    model_info: Final = _mapped_model_info(model, custom_llm_provider)
+    return model_info is not None and not model_info.get("supports_reasoning")
+
+
 def reasoning_off_level(model: str, custom_llm_provider: str | None, *, drop_params: bool) -> str | None:
     from litellm.utils import supports_none_reasoning_effort, supports_reasoning
 

tests/unit/llms/anthropic/messages/adapters/test_handler_reasoning_effort_normalization.py

@@ -737,3 +737,47 @@ class TestAnOpenAIChatRouteRequestGetsTheResponsesBridgeThinkingControls:
         )
 
         assert "reasoning_effort" not in sent
+
+
+_OFF_MAP_OPENAI_DEPLOYMENT: Final = {
+    "model": "openai/off-map-hybrid-reasoner",
+    "api_base": "https://compatible.example.test/v1",
+    "api_key": "fake-key",
+}
+
+
+class TestDisabledOnAnOffMapOpenAIDeployment:
+    @pytest.fixture(autouse=True)
+    def _httpx_transport(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+
+    @pytest.mark.asyncio
+    async def test_the_chat_route_sends_reasoning_effort_none(
+        self, respx_mock: respx.MockRouter, local_model_cost_map: None
+    ) -> None:
+        sent: Final = await _body_sent_upstream(
+            respx_mock,
+            _OFF_MAP_OPENAI_DEPLOYMENT,
+            {},
+            r"https://compatible\.example\.test/v1/chat/completions",
+            _CHAT_COMPLETION,
+            thinking={"type": "disabled"},
+            stop_sequences=["STOP"],
+        )
+
+        assert sent["reasoning_effort"] == "none"
+
+    @pytest.mark.asyncio
+    async def test_the_responses_bridge_sends_no_reasoning_its_transformation_would_refuse(
+        self, respx_mock: respx.MockRouter, local_model_cost_map: None
+    ) -> None:
+        sent: Final = await _body_sent_upstream(
+            respx_mock,
+            _OFF_MAP_OPENAI_DEPLOYMENT,
+            {},
+            r"https://compatible\.example\.test/v1/responses",
+            _RESPONSE,
+            thinking={"type": "disabled"},
+        )
+
+        assert "reasoning" not in sent