fix(passthrough): keep adaptive thinking/effort on native /v1/messages path

litellm/llms/openai_like/messages/transformation.py

@@ -80,7 +80,18 @@ class OpenAILikeAnthropicMessagesConfig(AnthropicMessagesConfig):
         on Anthropic-only ``cache_control`` extensions (``cache_control.ttl: 1h
         is not supported``), so unless the provider declares ttl support the
         hints are reduced to their portable ``{"type": ...}`` core.
+
+        The deployment speaks the Messages API natively, so caller-supplied
+        ``thinking``/``output_config``/``temperature`` are forwarded verbatim:
+        the base capability checks resolve the provider as "anthropic" and
+        would otherwise strip modern fields (e.g. adaptive thinking) from a
+        model the cost map does not know (see #40890). Pop them before the
+        base transform so no drop warning fires, restore afterwards.
         """
+        stashed: dict = {}
+        for _key in ("thinking", "output_config", "temperature"):
+            if _key in anthropic_messages_optional_request_params:
+                stashed[_key] = anthropic_messages_optional_request_params.pop(_key)
         request: Final = super().transform_anthropic_messages_request(
             model=model,
             messages=messages,
@@ -89,8 +100,11 @@ class OpenAILikeAnthropicMessagesConfig(AnthropicMessagesConfig):
             headers=headers,
         )
         if self.supports_cache_control_ttl():
+            request.update(stashed)
             return request
-        return normalize_cache_control_in_anthropic_payload(request)
+        normalized = normalize_cache_control_in_anthropic_payload(request)
+        normalized.update(stashed)
+        return normalized
 
     def get_complete_url(
         self,

tests/unit/llms/openai_like/messages/test_openai_like_anthropic_messages_transformation.py

@@ -518,3 +518,43 @@ def test_request_strips_ttl_only_where_the_messages_api_defines_cache_control(co
     assert tool_result["content"][0]["cache_control"] == {"type": "ephemeral"}
     assert payload["messages"][1]["content"][1]["cache_control"] == {"type": "ephemeral"}
     assert payload["messages"][2] == {"role": "user", "content": "a plain string message"}
+
+
+def test_passthrough_keeps_adaptive_thinking_and_effort(config, caplog):
+    """Regression #40890: native /v1/messages passthrough must forward
+    thinking/output_config/temperature verbatim, not strip via cost-map checks."""
+    import logging
+
+    optional_params = {
+        "max_tokens": 4096,
+        "thinking": {"type": "adaptive"},
+        "output_config": {"effort": "high"},
+        "temperature": 0.5,
+    }
+    with caplog.at_level(logging.WARNING):
+        payload = config.transform_anthropic_messages_request(
+            model="my-model",
+            messages=[{"role": "user", "content": "hi"}],
+            anthropic_messages_optional_request_params=dict(optional_params),
+            litellm_params=GenericLiteLLMParams(),
+            headers={},
+        )
+    assert payload["thinking"] == {"type": "adaptive"}
+    assert payload["output_config"] == {"effort": "high"}
+    assert payload["temperature"] == 0.5
+    assert "Dropping adaptive" not in caplog.text
+
+
+def test_passthrough_keeps_legacy_thinking_shape(config):
+    optional_params = {
+        "max_tokens": 4096,
+        "thinking": {"type": "enabled", "budget_tokens": 1024},
+    }
+    payload = config.transform_anthropic_messages_request(
+        model="my-model",
+        messages=[{"role": "user", "content": "hi"}],
+        anthropic_messages_optional_request_params=dict(optional_params),
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+    assert payload["thinking"] == {"type": "enabled", "budget_tokens": 1024}