refactor(responses): pass through extra usage fields generically

Avoid hard-coding provider-specific usage keys in shared Responses
utilities; forward any non-standard usage attributes onto chat Usage
for provider cost tracking (e.g. server_side_tool_usage_details).

litellm/responses/utils.py

@@ -1022,6 +1022,20 @@ class ResponsesAPIRequestUtils:
 
 
 class ResponseAPILoggingUtils:
+    # Standard Responses usage keys mapped explicitly below; extras pass through to Usage.
+    _RESPONSE_API_USAGE_MAPPED_KEYS = frozenset(
+        {
+            "input_tokens",
+            "output_tokens",
+            "total_tokens",
+            "input_tokens_details",
+            "output_tokens_details",
+            "input_token_details",
+            "output_token_details",
+            "cost",  # handled separately after Usage construction
+        }
+    )
+
     @staticmethod
     def _is_response_api_usage(usage: dict | ResponseAPIUsage) -> bool:
         """returns True if usage is from OpenAI Response API"""
@@ -1031,6 +1045,33 @@ class ResponseAPILoggingUtils:
             return True
         return False
 
+    @staticmethod
+    def _extra_fields_from_response_api_usage(
+        usage_input: dict | ResponseAPIUsage,
+        response_api_usage: ResponseAPIUsage,
+    ) -> dict[str, Any]:
+        """
+        Preserve provider/extension usage fields not part of the standard token mapping.
+
+        ResponseAPIUsage allows extra attributes; without forwarding them, the rebuilt
+        chat Usage would drop fields needed for provider-specific cost tracking.
+        """
+        extras: dict[str, Any] = {}
+        if isinstance(usage_input, dict):
+            for key, value in usage_input.items():
+                if key not in ResponseAPILoggingUtils._RESPONSE_API_USAGE_MAPPED_KEYS and value is not None:
+                    extras[key] = value
+            return extras
+
+        model_extra = getattr(response_api_usage, "model_extra", None) or getattr(
+            response_api_usage, "__pydantic_extra__", None
+        )
+        if isinstance(model_extra, dict):
+            for key, value in model_extra.items():
+                if key not in ResponseAPILoggingUtils._RESPONSE_API_USAGE_MAPPED_KEYS and value is not None:
+                    extras[key] = value
+        return extras
+
     @staticmethod
     def _transform_response_api_usage_to_chat_usage(
         usage_input: dict | ResponseAPIUsage | None,
@@ -1089,17 +1130,10 @@ class ResponseAPILoggingUtils:
                 audio_tokens=getattr(output_tokens_details, "audio_tokens", None),
             )
 
-        usage_kwargs: dict[str, Any] = {}
-        # Keep xAI tool billing fields; dropped if we only pass token fields below.
-        if isinstance(usage_input, dict):
-            if usage_input.get("server_side_tool_usage_details") is not None:
-                usage_kwargs["server_side_tool_usage_details"] = usage_input["server_side_tool_usage_details"]
-        else:
-            details = getattr(
-                response_api_usage, "server_side_tool_usage_details", None
-            )
-            if details is not None:
-                usage_kwargs["server_side_tool_usage_details"] = details
+        usage_kwargs: Final = ResponseAPILoggingUtils._extra_fields_from_response_api_usage(
+            usage_input=usage_input,
+            response_api_usage=response_api_usage,
+        )
 
         chat_usage: Final = Usage(
             prompt_tokens=prompt_tokens,

tests/test_litellm/responses/test_responses_utils.py

@@ -278,6 +278,23 @@ class TestResponseAPILoggingUtils:
             and result.prompt_tokens_details.cached_tokens == 2
         )
 
+    def test_transform_response_api_usage_preserves_extra_usage_fields(self):
+        """Non-standard usage keys pass through for provider cost tracking."""
+        usage = {
+            "input_tokens": 10,
+            "output_tokens": 5,
+            "total_tokens": 15,
+            "server_side_tool_usage_details": {"web_search_calls": 2},
+            "num_server_side_tools_used": 2,
+        }
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
+            usage
+        )
+        assert getattr(result, "server_side_tool_usage_details", None) == {
+            "web_search_calls": 2
+        }
+        assert getattr(result, "num_server_side_tools_used", None) == 2
+
     def test_transform_response_api_usage_with_none_values(self):
         """Test transformation handles None values properly"""
         # Setup