fix(logging): price terminal Responses stream events from their inner response

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/litellm_core_utils/litellm_logging.py

@@ -1757,13 +1757,19 @@ class Logging(LiteLLMLoggingBaseClass):
         if transformed_result is not None:
             result = transformed_result
 
-        result_hidden_params: Final = getattr(result, "_hidden_params", None) or MappingProxyType({})
+        priced_result: Final = (
+            result.response
+            if isinstance(result, (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent))
+            else result
+        )
+
+        result_hidden_params: Final = getattr(priced_result, "_hidden_params", None) or MappingProxyType({})
         result_additional_headers: Final = (
             result_hidden_params.get("additional_headers")
             if isinstance(result_hidden_params, dict)
             else getattr(result_hidden_params, "additional_headers", None)
         )
-        if isinstance(result, (BaseModel, HttpxBinaryResponseContent)) and hasattr(result, "_hidden_params"):
+        if isinstance(priced_result, (BaseModel, HttpxBinaryResponseContent)) and hasattr(priced_result, "_hidden_params"):
             hidden_params: Final = result_hidden_params
             if (
                 "response_cost" in hidden_params and hidden_params["response_cost"] is not None
@@ -1799,7 +1805,7 @@ class Logging(LiteLLMLoggingBaseClass):
 
         try:
             response_cost_calculator_kwargs: Final = {
-                "response_object": result,
+                "response_object": priced_result,
                 "model": litellm_model_name or self.model,
                 "cache_hit": cache_hit,
                 "custom_llm_provider": self.model_call_details.get("custom_llm_provider", None),

litellm/responses/streaming_iterator.py

@@ -801,13 +801,8 @@ class BaseResponsesAPIStreamingIterator:
                 request_payload["litellm_params"] = {}
 
         try:
-            response_obj: Final = self._get_completed_response_object()
             update_response_metadata(
-                result=(
-                    type(response_obj).model_validate(response_obj.model_dump())
-                    if response_obj is not None
-                    else self.completed_response
-                ),
+                result=self.completed_response,
                 logging_obj=self.logging_obj,
                 model=self.model,
                 kwargs=request_payload,

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -7467,3 +7467,26 @@ def test_get_assembled_streaming_response_without_usage_cost_leaves_pricing_to_t
     assert "additional_headers" not in assembled._hidden_params
     price_map_cost = logging_obj._response_cost_calculator(result=assembled)
     assert price_map_cost is not None and 0 < price_map_cost != 0.0042
+
+
+def test_response_cost_calculator_prices_terminal_responses_event_from_its_response():
+    """A terminal Responses stream event carries no usage itself; pricing must unwrap
+    it so the stored cost_breakdown is not overwritten with zeros."""
+    logging_obj: Final = _responses_stream_logging_obj()
+    inner_response: Final = ResponsesAPIResponse(
+        id="resp-priced",
+        created_at=1,
+        object="response",
+        status="completed",
+        model="gpt-4o-mini",
+        output=[],
+        usage=ResponseAPIUsage(input_tokens=1840, output_tokens=412, total_tokens=2252),
+    )
+    event: Final = ResponseCompletedEvent(type="response.completed", response=inner_response)
+
+    event_cost: Final = logging_obj._response_cost_calculator(result=event)
+    inner_cost: Final = logging_obj._response_cost_calculator(result=inner_response)
+
+    assert event_cost is not None and event_cost > 0
+    assert event_cost == inner_cost
+    assert logging_obj.cost_breakdown["input_cost"] is not None and logging_obj.cost_breakdown["input_cost"] > 0

tests/test_litellm/responses/test_streaming_iterator.py

@@ -342,46 +342,6 @@ def test_run_post_success_hooks_does_not_report_generation_time_as_overhead():
     assert "litellm_overhead_time_ms" not in iterator.completed_response._hidden_params
 
 
-def test_run_post_success_hooks_prices_the_completed_response_not_the_event():
-    """The terminal ResponseCompletedEvent carries no usage, so pricing it recomputes
-    cost as 0 and clobbers the cost_breakdown the stream already stored."""
-    inner_response: Final = ResponsesAPIResponse(
-        id="resp_pricing",
-        created_at=0,
-        status="completed",
-        model="gpt-4o-mini",
-        object="response",
-        output=[],
-        usage=ResponseAPIUsage(input_tokens=1840, output_tokens=412, total_tokens=2252),
-    )
-    event: Final = ResponseCompletedEvent(
-        type=ResponsesAPIStreamEvents.RESPONSE_COMPLETED,
-        response=inner_response,
-    )
-
-    logging_obj = _logging_obj_stub()
-    priced_results: Final[list[object]] = []
-
-    def _cost_calculator(*, result, **_kwargs):
-        priced_results.append(result)
-        return 0.0156
-
-    logging_obj._response_cost_calculator.side_effect = _cost_calculator
-
-    iterator = _make_iterator(sse_events=[], logging_obj=logging_obj)
-    iterator.completed_response = event
-    iterator.start_time = datetime(2025, 1, 1, 0, 0, 0)
-
-    iterator._run_post_success_hooks(datetime(2025, 1, 1, 0, 0, 10))
-
-    assert priced_results and all(
-        isinstance(result, ResponsesAPIResponse) and result is not event and result.usage is not None
-        for result in priced_results
-    )
-    assert priced_results[0]._hidden_params["response_cost"] == 0.0156
-    assert event.response._hidden_params == {}
-
-
 def _mock_config_with_completed_response(response: ResponsesAPIResponse) -> Mock:
     mock_config = Mock(spec=BaseResponsesAPIConfig)