fix(azure_ai): let the caller's output_config from extra_body win over the legacy thinking upgrade

litellm/llms/azure_ai/anthropic/transformation.py

@@ -17,13 +17,14 @@ def _promote_extra_body_to_optional_params(optional_params: dict) -> None:
     ``output_config`` get auto-routed into ``extra_body`` by
     ``add_provider_specific_params_to_optional_params``. For the Azure→Anthropic
     route those keys must reach the request body and be validated, so promote
-    them. ``setdefault`` keeps explicit top-level values authoritative.
+    them. The caller's values overwrite mapped top-level duplicates, matching
+    the native ``anthropic`` provider, where the same passthrough lands on
+    top-level ``optional_params`` after mapping.
     """
     extra_body: Final = optional_params.get("extra_body")
     if not isinstance(extra_body, dict) or not extra_body:
         return
-    for k, v in extra_body.items():
-        optional_params.setdefault(k, v)
+    optional_params.update(extra_body)
     optional_params.pop("extra_body", None)
 
 

tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_transformation.py

@@ -362,8 +362,8 @@ class TestAzureAnthropicConfig:
             )
         assert "xhigh" in str(exc_info.value)
 
-    def test_extra_body_promotion_does_not_clobber_top_level(self):
-        """Top-level ``optional_params`` wins over duplicates in ``extra_body``."""
+    def test_extra_body_promotion_overrides_mapped_top_level(self):
+        """The caller's ``extra_body`` wins over a mapped top-level duplicate, like the native ``anthropic`` passthrough."""
         config = AzureAnthropicConfig()
 
         messages = [{"role": "user", "content": "Hello"}]
@@ -383,7 +383,31 @@ class TestAzureAnthropicConfig:
             headers=headers,
         )
 
-        assert result["output_config"] == {"effort": "low"}
+        assert result["output_config"] == {"effort": "high"}
+
+    def test_legacy_thinking_upgrade_keeps_caller_effort_from_extra_body(self, local_model_cost_map):
+        config = AzureAnthropicConfig()
+
+        mapped = config.map_openai_params(
+            non_default_params={"thinking": {"type": "enabled", "budget_tokens": 1024}, "max_tokens": 100},
+            optional_params={},
+            model="claude-opus-4-8",
+            drop_params=False,
+        )
+        assert mapped["thinking"] == {"type": "adaptive"}
+        assert mapped["output_config"] == {"effort": "low"}
+
+        result = config.transform_request(
+            model="claude-opus-4-8",
+            messages=[{"role": "user", "content": "Hello"}],
+            optional_params={**mapped, "extra_body": {"output_config": {"effort": "high"}}},
+            litellm_params={"api_key": "test-key"},
+            headers={"api-key": "test-key", "anthropic-version": "2023-06-01"},
+        )
+
+        assert result["thinking"] == {"type": "adaptive"}
+        assert result["output_config"] == {"effort": "high"}
+        assert "extra_body" not in result
 
     def test_context_management_mixed_edits_beta_headers(self):
         """Test that context_management with both compact and other edits adds both beta headers"""