fix(bedrock): preserve cache_control for ARN models in /v1/messages adapter (#29823)

* fix(bedrock): preserve cache_control for ARN models in /v1/messages adapter

Bedrock Application Inference Profile ARNs contain neither "anthropic" nor
"claude", so is_anthropic_claude_model could not detect them and the
/v1/messages adapter silently dropped cache_control during the Anthropic to
OpenAI translation. Prompt caching never activated for these models, while the
same profile cached correctly through /v1/chat/completions.

Add an is_bedrock_arn_model check scoped to _add_cache_control_if_applicable so
cache_control is preserved for ARN-based models without broadening the shared
is_anthropic_claude_model helper, which also drives thinking translation.

Fixes #26625

* refactor(bedrock): match :bedrock: ARN service field in is_bedrock_arn_model

Tighten the ARN detection so it pins "bedrock" to the colon-delimited service
field of the ARN rather than matching the substring anywhere. This avoids a
false positive for another service's ARN whose resource name merely contains
"bedrock" (e.g. arn:aws:sagemaker:...:endpoint/my-bedrock-transcriber).

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

@@ -332,7 +332,14 @@ class LiteLLMAnthropicMessagesAdapter:
             if isinstance(source, dict)
             else getattr(source, "cache_control", None)
         )
-        if cache_control and model and self.is_anthropic_claude_model(model):
+        if (
+            cache_control
+            and model
+            and (
+                self.is_anthropic_claude_model(model)
+                or self.is_bedrock_arn_model(model)
+            )
+        ):
             # TypedDict objects support dict operations at runtime
             # Use type ignore consistent with codebase pattern (see anthropic/chat/transformation.py:432)
             if isinstance(target, dict):
@@ -752,6 +759,20 @@ class LiteLLMAnthropicMessagesAdapter:
         model_lower = model.lower()
         return "anthropic" in model_lower or "claude" in model_lower
 
+    @staticmethod
+    def is_bedrock_arn_model(model: str) -> bool:
+        """
+        Check if the model string is a Bedrock ARN, such as an Application
+        Inference Profile (e.g. arn:aws:bedrock:us-east-1:123:application-inference-profile/id).
+
+        These ARNs contain neither "anthropic" nor "claude", so is_anthropic_claude_model
+        cannot identify them even though, on the /v1/messages endpoint, they point at Claude.
+        Match ":bedrock:" in the ARN service field so another service's ARN that merely names
+        bedrock in a resource (arn:aws:sagemaker:.../my-bedrock-endpoint) is not matched.
+        """
+        model_lower = model.lower()
+        return "arn:" in model_lower and ":bedrock:" in model_lower
+
     @staticmethod
     def translate_thinking_for_model(
         thinking: Dict[str, Any],

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -1295,6 +1295,12 @@ CACHE_CONTROL_BEDROCK_CONVERSE_MODEL = (
     "bedrock/converse/global.anthropic.claude-opus-4-5-20251101-v1:0"
 )
 CACHE_CONTROL_NON_ANTHROPIC_MODEL = "gpt-4"
+# Bedrock Application Inference Profile ARN: the string contains neither
+# "anthropic" nor "claude", so the model can only be recognized via its ARN shape
+CACHE_CONTROL_BEDROCK_ARN_MODEL = (
+    "bedrock/converse/arn:aws:bedrock:us-east-1:123456789012:"
+    "application-inference-profile/abcdef123456"
+)
 
 
 def test_should_add_cache_control_for_anthropic_model():
@@ -1411,6 +1417,68 @@ def test_cache_control_not_preserved_for_non_claude_model():
     assert "cache_control" not in result[0]["content"][0]
 
 
+@pytest.mark.parametrize(
+    "model, expected",
+    [
+        (CACHE_CONTROL_BEDROCK_ARN_MODEL, True),
+        (
+            "arn:aws-us-gov:bedrock:us-gov-west-1:123:application-inference-profile/x",
+            True,
+        ),
+        ("bedrock/amazon.titan-text-express-v1", False),
+        ("arn:aws:sagemaker:us-east-1:123:endpoint/my-endpoint", False),
+        ("arn:aws:sagemaker:us-east-1:123:endpoint/my-bedrock-transcriber", False),
+        (CACHE_CONTROL_NON_ANTHROPIC_MODEL, False),
+    ],
+)
+def test_is_bedrock_arn_model(model, expected):
+    """is_bedrock_arn_model requires an ARN with bedrock in the service field, not just anywhere."""
+    assert LiteLLMAnthropicMessagesAdapter.is_bedrock_arn_model(model) is expected
+
+
+def test_cache_control_preserved_for_bedrock_arn_inference_profile():
+    """
+    Regression for https://github.com/BerriAI/litellm/issues/26625
+
+    Bedrock Application Inference Profile ARNs hide the underlying Claude model
+    name, so cache_control must still be preserved through the /v1/messages adapter.
+    """
+    anthropic_messages = [
+        AnthropicMessagesUserMessageParam(
+            role="user",
+            content=[
+                {
+                    "type": "text",
+                    "text": "This is cached content",
+                    "cache_control": {"type": "ephemeral"},
+                }
+            ],
+        )
+    ]
+
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    result = adapter.translate_anthropic_messages_to_openai(
+        messages=anthropic_messages, model=CACHE_CONTROL_BEDROCK_ARN_MODEL
+    )
+
+    assert len(result) == 1
+    assert result[0]["content"][0]["cache_control"] == {"type": "ephemeral"}
+
+
+def test_cache_control_fix_does_not_broaden_claude_detection():
+    """
+    The cache_control fix is scoped to _add_cache_control_if_applicable; it must not
+    make is_anthropic_claude_model treat ARN profiles as Claude, which would route
+    thinking params through unmodified and break non-Claude Bedrock profiles.
+    """
+    assert (
+        LiteLLMAnthropicMessagesAdapter.is_anthropic_claude_model(
+            CACHE_CONTROL_BEDROCK_ARN_MODEL
+        )
+        is False
+    )
+
+
 def test_cache_control_preserved_in_image_content_for_claude():
     """Cache control should be preserved in image content for Claude models."""
     anthropic_messages = [