fix: preserve cache_creation fields in responses API usage transform

The _transform_chat_completion_usage_to_responses_usage method drops
Anthropic cache creation fields (cache_creation_input_tokens,
cache_read_input_tokens, cache_creation.ephemeral_5m/1h_input_tokens)
when converting chat-completion Usage to ResponseAPIUsage. This causes
the inference proxy to see zero cache creation tokens, resulting in
unbilled cache creation on /v1/responses.

Fix: use setattr to add these fields as extras on ResponseAPIUsage
(which supports extra fields via BaseLiteLLMOpenAIResponseObject's
extra="allow" config), so they survive serialization and can be
extracted downstream.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: lukesorvikDO

litellm/responses/litellm_completion_transformation/transformation.py

@@ -2147,6 +2147,26 @@ def _transform_chat_completion_usage_to_responses_usage(
                     **output_details_dict
                 )
 
+        # Preserve Anthropic cache creation fields for downstream billing extraction.
+        # These are not part of the OpenAI ResponseAPIUsage schema but
+        # BaseLiteLLMOpenAIResponseObject has extra="allow", so setattr ensures
+        # they survive serialization and can be extracted by the inference proxy
+        # before the lossy api.Usage unmarshal.
+        if hasattr(usage, "cache_creation_input_tokens") and usage.cache_creation_input_tokens:
+            setattr(response_usage, "cache_creation_input_tokens", usage.cache_creation_input_tokens)
+        if hasattr(usage, "cache_read_input_tokens") and usage.cache_read_input_tokens:
+            setattr(response_usage, "cache_read_input_tokens", usage.cache_read_input_tokens)
+        if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details is not None:
+            ptd = usage.prompt_tokens_details
+            if hasattr(ptd, "cache_creation_token_details") and ptd.cache_creation_token_details is not None:
+                cache_creation_dict: Dict[str, int] = {}
+                if ptd.cache_creation_token_details.ephemeral_5m_input_tokens is not None:
+                    cache_creation_dict["ephemeral_5m_input_tokens"] = ptd.cache_creation_token_details.ephemeral_5m_input_tokens
+                if ptd.cache_creation_token_details.ephemeral_1h_input_tokens is not None:
+                    cache_creation_dict["ephemeral_1h_input_tokens"] = ptd.cache_creation_token_details.ephemeral_1h_input_tokens
+                if cache_creation_dict:
+                    setattr(response_usage, "cache_creation", cache_creation_dict)
+
         return response_usage
 
     @staticmethod