fix(bedrock): forward output_config effort for application inference profile ARNs

litellm/llms/bedrock/chat/converse_transformation.py

@@ -80,6 +80,7 @@ from ..common_utils import (
     bedrock_converse_supports_parallel_tool_use_config,
     get_anthropic_beta_from_headers,
     get_bedrock_tool_name,
+    is_bedrock_application_inference_profile_arn,
     is_claude_4_5_on_bedrock,
     normalize_bedrock_opus_output_config_effort,
 )
@@ -1318,7 +1319,12 @@ class AmazonConverseConfig(BaseConfig):
         additional_request_params = filter_exceptions_from_params(additional_request_params)
 
         if anthropic_output_config is not None and isinstance(anthropic_output_config, dict):
-            if base_model.startswith("anthropic"):
+            # Application inference profile ARNs hide the underlying model, so the
+            # effort ceiling and capability gates below cannot run; forward
+            # verbatim (like ``thinking``) and let Bedrock enforce.
+            if is_bedrock_application_inference_profile_arn(model):
+                additional_request_params["output_config"] = anthropic_output_config
+            elif base_model.startswith("anthropic"):
                 if litellm.drop_params is True and not AnthropicConfig._model_supports_effort_param(model, "bedrock"):
                     litellm.verbose_logger.warning(
                         DROP_UNSUPPORTED_OUTPUT_CONFIG_WARNING,

tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py

@@ -410,6 +410,33 @@ def test_output_config_supported_param_for_arn_models_converse():
     assert "output_config" in config.get_supported_openai_params(arn_model)
 
 
+def test_output_config_effort_forwarded_for_application_inference_profile_arn():
+    """Regression: opaque application inference profile ARNs cannot resolve a
+    base model, so the anthropic-only serialization gate dropped ``output_config``
+    while still sending ``thinking``: adaptive thinking with no effort tier, and
+    Bedrock streams zero ``reasoningContent`` blocks. The effort must be forwarded
+    verbatim (ceilings and capability gates are unknowable behind the alias) for
+    Bedrock to enforce."""
+    config = AmazonConverseConfig()
+    arn_model = "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abcdef123456"
+
+    result = config._transform_request(
+        model=arn_model,
+        messages=[{"role": "user", "content": "hi"}],
+        optional_params={
+            "maxTokens": 256,
+            "thinking": {"type": "adaptive"},
+            "output_config": {"effort": "max"},
+        },
+        litellm_params={},
+        headers={},
+    )
+
+    additional = result.get("additionalModelRequestFields", {})
+    assert additional.get("thinking") == {"type": "adaptive"}
+    assert additional.get("output_config") == {"effort": "max"}
+
+
 def test_output_config_format_translated_to_native_output_config_converse():
     """``output_config.format`` becomes Bedrock ``outputConfig`` and is not forwarded raw."""
     config = AmazonConverseConfig()