[Bug fix] - Fix /messages fallback from Anthropic API -> Bedrock API (#13946)

* use helper get_provider_specific_headers

* fix get_provider_specific_headers

* test_anthropic_messages_fallbacks

* bedrock/us.anthropic.claude-sonnet-4

* fix: get_provider_specific_headers

* TestProviderSpecificHeaderUtils

* test_anthropic_messages_fallbacks

Author: ishaan-jaff

litellm/litellm_core_utils/get_provider_specific_headers.py

@@ -0,0 +1,23 @@
+from typing import Dict, Optional
+
+from litellm.types.utils import ProviderSpecificHeader
+
+
+class ProviderSpecificHeaderUtils:
+    @staticmethod
+    def get_provider_specific_headers(
+        provider_specific_header: Optional[ProviderSpecificHeader],
+        custom_llm_provider: Optional[str],
+    ) -> Dict:
+        """
+        Get the provider specific headers for the given custom llm provider
+
+        Returns:
+            Optional[Dict]: The provider specific headers for the given custom llm provider
+        """
+        if (
+            provider_specific_header is not None
+            and provider_specific_header.get("custom_llm_provider") == custom_llm_provider
+        ):
+            return provider_specific_header.get("extra_headers", {})
+        return {}

litellm/llms/custom_httpx/llm_http_handler.py

@@ -1257,6 +1257,10 @@ async def async_anthropic_messages_handler(
         stream: Optional[bool] = False,
         kwargs: Optional[Dict[str, Any]] = None,
     ) -> Union[AnthropicMessagesResponse, AsyncIterator]:
+        from litellm.litellm_core_utils.get_provider_specific_headers import (
+            ProviderSpecificHeaderUtils,
+        )
+
         if client is None or not isinstance(client, AsyncHTTPHandler):
             async_httpx_client = get_async_httpx_client(
                 llm_provider=litellm.LlmProviders.ANTHROPIC
@@ -1270,10 +1274,9 @@ async def async_anthropic_messages_handler(
             Optional[litellm.types.utils.ProviderSpecificHeader],
             kwargs.get("provider_specific_header", None),
         )
-        extra_headers = (
-            provider_specific_header.get("extra_headers", {})
-            if provider_specific_header
-            else {}
+        extra_headers = ProviderSpecificHeaderUtils.get_provider_specific_headers(
+            provider_specific_header=provider_specific_header,
+            custom_llm_provider=custom_llm_provider,
         )
         (
             headers,

litellm/main.py

@@ -61,6 +61,9 @@
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.audio_utils.utils import get_audio_file_for_health_check
 from litellm.litellm_core_utils.dd_tracing import tracer
+from litellm.litellm_core_utils.get_provider_specific_headers import (
+    ProviderSpecificHeaderUtils,
+)
 from litellm.litellm_core_utils.health_check_utils import (
     _create_health_check_response,
     _filter_model_params,
@@ -1107,11 +1110,11 @@ def completion(  # type: ignore # noqa: PLR0915
             api_key=api_key,
         )
 
-        if (
-            provider_specific_header is not None
-            and provider_specific_header["custom_llm_provider"] == custom_llm_provider
-        ):
-            headers.update(provider_specific_header["extra_headers"])
+        if provider_specific_header is not None:
+            headers.update(ProviderSpecificHeaderUtils.get_provider_specific_headers(
+                provider_specific_header=provider_specific_header,
+                custom_llm_provider=custom_llm_provider,
+            ))
 
         if model_response is not None and hasattr(model_response, "_hidden_params"):
             model_response._hidden_params["custom_llm_provider"] = custom_llm_provider

litellm/proxy/proxy_config.yaml

@@ -1,17 +1,23 @@
 model_list:
-  - model_name: bedrock/converse/us.anthropic.claude-sonnet-4-20250514-v1:0
+  - model_name: anthropic/*
     litellm_params:
-      model: bedrock/converse/us.anthropic.claude-sonnet-4-20250514-v1:0
+      model: anthropic/*
+      api_key: os.environ/OPENAI_API_KEY_IJ
   - model_name: bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0
     litellm_params:
       model: bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0
+  - model_name: bedrock/converse/us.anthropic.claude-sonnet-4-20250514-v1:0
+    litellm_params:
+      model: bedrock/converse/us.anthropic.claude-sonnet-4-20250514-v1:0
+
+router_settings:
+  fallbacks: [
+    {"anthropic/claude-opus-4-20250514": 
+      {
+        "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"
+      }
+    }
+  ]
 litellm_settings:
   callbacks: ["datadog_llm_observability"]
 
-guardrails:
-  - guardrail_name: "bedrock-pre-guard"
-    litellm_params:
-      guardrail: bedrock  # supported values: "aporia", "bedrock", "lakera"
-      mode: "during_call"
-      guardrailIdentifier: ff6ujrregl1q
-      guardrailVersion: "DRAFT"

tests/pass_through_unit_tests/test_anthropic_messages_passthrough.py

@@ -273,6 +273,58 @@ async def test_anthropic_messages_litellm_router_routing_strategy():
     print(f"Non-streaming response: {json.dumps(response, indent=2)}")
     return response
 
+@pytest.mark.asyncio
+async def test_anthropic_messages_fallbacks():
+    """
+    E2E test the anthropic_messages fallbacks from Anthropic API to Bedrock
+    """
+    litellm._turn_on_debug()
+    router = Router(
+        model_list=[
+            {
+                "model_name": "anthropic/claude-opus-4-20250514",
+                "litellm_params": {
+                    "model": "anthropic/claude-opus-4-20250514",
+                    "api_key": "bad-key",
+                },
+            },
+            {
+                "model_name": "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0",
+                "litellm_params": {
+                    "model": "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0",
+                },
+            }
+        ],
+        fallbacks=[
+            {
+                "anthropic/claude-opus-4-20250514": 
+                ["bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"]
+            }
+        ]
+    )
+
+    # Set up test parameters
+    messages = [{"role": "user", "content": "Hello, can you tell me a short joke?"}]
+
+    # Call the handler
+    response = await router.aanthropic_messages(
+        messages=messages,
+        model="anthropic/claude-opus-4-20250514",
+        max_tokens=100,
+        metadata={
+            "user_id": "hello",
+        },
+    )
+
+    # Verify response
+    assert "id" in response
+    assert "content" in response
+    assert "model" in response
+    assert response["role"] == "assistant"
+
+    print(f"Non-streaming response: {json.dumps(response, indent=2)}")
+    return response
+
 
 @pytest.mark.asyncio
 async def test_anthropic_messages_litellm_router_latency_metadata_tracking():

tests/test_litellm/litellm_core_utils/test_provider_specific_headers.py

@@ -0,0 +1,43 @@
+import pytest
+
+from litellm.litellm_core_utils.get_provider_specific_headers import (
+    ProviderSpecificHeaderUtils,
+)
+from litellm.types.utils import ProviderSpecificHeader
+
+
+class TestProviderSpecificHeaderUtils:
+    def test_get_provider_specific_headers_matching_provider(self):
+        """Test that the method returns extra_headers when custom_llm_provider matches."""
+        provider_specific_header: ProviderSpecificHeader = {
+            "custom_llm_provider": "openai",
+            "extra_headers": {"Authorization": "Bearer token123", "Custom-Header": "value"}
+        }
+        custom_llm_provider = "openai"
+        
+        result = ProviderSpecificHeaderUtils.get_provider_specific_headers(
+            provider_specific_header, custom_llm_provider
+        )
+        
+        expected = {"Authorization": "Bearer token123", "Custom-Header": "value"}
+        assert result == expected
+
+    def test_get_provider_specific_headers_no_match_or_none(self):
+        """Test that the method returns empty dict when provider doesn't match or is None."""
+        # Test case 1: Provider doesn't match
+        provider_specific_header: ProviderSpecificHeader = {
+            "custom_llm_provider": "anthropic",
+            "extra_headers": {"Authorization": "Bearer token123"}
+        }
+        custom_llm_provider = "openai"
+        
+        result = ProviderSpecificHeaderUtils.get_provider_specific_headers(
+            provider_specific_header, custom_llm_provider
+        )
+        assert result == {}
+        
+        # Test case 2: provider_specific_header is None
+        result = ProviderSpecificHeaderUtils.get_provider_specific_headers(
+            None, "openai"
+        )
+        assert result == {}