fix(bedrock): send an application inference profile's effort when its base_model is set

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

@@ -646,6 +646,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             request_data["output_config"] = extra_kwargs["output_config"]
 
         custom_llm_provider: Final = extra_kwargs.get("custom_llm_provider")
+        base_model: Final = extra_kwargs.get("base_model")
         (
             openai_request,
             tool_name_mapping,
@@ -653,6 +654,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             request_data,
             custom_llm_provider=custom_llm_provider if isinstance(custom_llm_provider, str) else None,
             drop_params=drops_unsupported_params(extra_kwargs.get("drop_params")),
+            base_model=base_model if isinstance(base_model, str) else None,
         )
 
         if openai_request is None:

litellm/llms/anthropic/messages/adapters/transformation.py

@@ -16,7 +16,7 @@ from litellm.llms.anthropic.messages.utils import (
     json_copy_or_original,
     prompt_cache_key_from_user_id,
 )
-from litellm.llms.bedrock.common_utils import is_bedrock_application_inference_profile_arn
+from litellm.llms.bedrock.common_utils import bedrock_capability_model, is_bedrock_application_inference_profile_arn
 from litellm.responses.utils import auto_reasoning_summary
 
 # OpenAI has a 64-character limit for function/tool names
@@ -256,7 +256,12 @@ class AnthropicAdapter:
         return result
 
     def translate_completion_input_params_with_tool_mapping(
-        self, kwargs, *, custom_llm_provider: str | None = None, drop_params: bool = True
+        self,
+        kwargs,
+        *,
+        custom_llm_provider: str | None = None,
+        drop_params: bool = True,
+        base_model: str | None = None,
     ) -> tuple[ChatCompletionRequest | None, dict[str, str]]:
         """
         Translate Anthropic request params to OpenAI format, returning tool name mapping.
@@ -291,6 +296,7 @@ class AnthropicAdapter:
             anthropic_message_request=request_body,
             custom_llm_provider=custom_llm_provider,
             drop_params=drop_params,
+            base_model=base_model,
         )
 
         return translated_body, tool_name_mapping
@@ -369,8 +375,8 @@ _BlockT: Final = TypeVar("_BlockT", bound=Mapping[str, object])
 _CLAUDE_OVER_REASONING_EFFORT_PROVIDERS: Final = frozenset({LlmProviders.KYMA.value})
 UNRESOLVED_ARN_EFFORT_WITHOUT_THINKING_WARNING: Final = (
     "Dropping `output_config` for model=%s under drop_params: an application inference profile ARN does not name "
-    "its model, so LiteLLM cannot tell whether it takes an effort without a `thinking` block. Turn off drop_params "
-    "on the deployment to send it."
+    "its model, so LiteLLM cannot tell whether it takes an effort without a `thinking` block. Set `base_model` on "
+    "the deployment (for example `us.anthropic.claude-opus-4-8`) to send it."
 )
 
 
@@ -1210,6 +1216,7 @@ class LiteLLMAnthropicMessagesAdapter:
         anthropic_message_request: AnthropicMessagesRequest,
         new_kwargs: ChatCompletionRequest,
         drop_params: bool,
+        base_model: str | None,
     ) -> None:
         """Forward the effort-bearing part of ``output_config`` to a Bedrock Claude target.
 
@@ -1231,7 +1238,7 @@ class LiteLLMAnthropicMessagesAdapter:
         if (
             drop_params
             and not anthropic_message_request.get("thinking")
-            and is_bedrock_application_inference_profile_arn(model)
+            and is_bedrock_application_inference_profile_arn(bedrock_capability_model(model, base_model))
         ):
             litellm.verbose_logger.warning(UNRESOLVED_ARN_EFFORT_WITHOUT_THINKING_WARNING, model)
             return
@@ -1283,6 +1290,7 @@ class LiteLLMAnthropicMessagesAdapter:
         custom_llm_provider: str | None = None,
         preserve_midturn_system: bool = False,
         drop_params: bool = True,
+        base_model: str | None = None,
     ) -> tuple[ChatCompletionRequest, dict[str, str]]:
         """
         This is used by the beta Anthropic Adapter, for translating anthropic `/v1/messages` requests to the openai format.
@@ -1341,6 +1349,7 @@ class LiteLLMAnthropicMessagesAdapter:
             anthropic_message_request=anthropic_message_request,
             new_kwargs=new_kwargs,
             drop_params=drop_params,
+            base_model=base_model,
         )
         ## CONVERT STOP_SEQUENCES
         self._translate_stop_sequences_to_openai(

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

@@ -5560,17 +5560,19 @@ _APPLICATION_INFERENCE_PROFILE: Final = (
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
-    ("thinking", "drop_params", "forwards_effort"),
+    ("thinking", "drop_params", "base_model", "forwards_effort"),
     [
-        pytest.param(None, False, True, id="no_thinking"),
-        pytest.param(None, True, False, id="no_thinking_drop_params"),
-        pytest.param({"type": "enabled", "budget_tokens": 1024}, True, True, id="thinking_drop_params"),
+        pytest.param(None, False, None, True, id="no_thinking"),
+        pytest.param(None, True, None, False, id="no_thinking_drop_params"),
+        pytest.param({"type": "enabled", "budget_tokens": 1024}, True, None, True, id="thinking_drop_params"),
+        pytest.param(None, True, _CONVERSE_EFFORT_MODEL, True, id="no_thinking_drop_params_base_model"),
     ],
 )
-async def test_bedrock_application_inference_profile_loses_the_effort_only_without_thinking_under_drop_params(
+async def test_bedrock_application_inference_profile_loses_the_effort_only_when_unresolved_without_thinking(
     _converse_route: respx.Route,
     thinking: Mapping[str, object] | None,
     drop_params: bool,
+    base_model: str | None,
     forwards_effort: bool,
 ) -> None:
     await litellm.anthropic.messages.acreate(
@@ -5580,6 +5582,7 @@ async def test_bedrock_application_inference_profile_loses_the_effort_only_witho
         output_config={"effort": "high", "format": _CITY_FORMAT},
         drop_params=drop_params,
         **({"thinking": thinking} if thinking else {}),
+        **({"base_model": base_model} if base_model else {}),
     )
 
     sent: Final = _sent_converse_fields(_converse_route)