fix(responses): keep the usage estimate best-effort when token counting raises

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

litellm/responses/streaming_iterator.py

@@ -440,7 +440,7 @@ class BaseResponsesAPIStreamingIterator:
                         and _response_obj is not None
                         and _response_obj.usage is None
                     ):
-                        _response_obj.usage = _estimate_usage_from_text(
+                        _response_obj.usage = _estimate_usage_safely(
                             self.model or "",
                             self.request_data.get("input"),
                             self.request_data,
@@ -1381,6 +1381,24 @@ def _estimate_usage_from_text(
     )
 
 
+def _estimate_usage_safely(
+    model: str,
+    request_input: object,
+    responses_api_request: Mapping[str, object],
+    generated_text: str,
+) -> ResponseAPIUsage | None:
+    try:
+        return _estimate_usage_from_text(
+            model=model,
+            request_input=request_input,
+            responses_api_request=responses_api_request,
+            generated_text=generated_text,
+        )
+    except Exception as e:
+        verbose_logger.debug("Could not estimate usage from stream text, billing $0: %s", e)
+        return None
+
+
 def _stamp_responses_usage_cost(
     response_obj: ResponsesAPIResponse | None, logging_obj: LiteLLMLoggingObj | None
 ) -> None:

tests/test_litellm/responses/test_streaming_iterator.py

@@ -792,3 +792,27 @@ async def test_completed_event_without_usage_counts_multimodal_input_as_messages
     usage = iterator.completed_response.response.usage
     assert usage is not None
     assert usage.input_tokens < json_count / 2
+
+
+@pytest.mark.asyncio
+async def test_completed_event_survives_a_failing_usage_estimate():
+    """A raising token_counter must not break a stream that previously completed:
+    the estimate is best-effort and falls back to usage None."""
+    response = _responses_api_response_without_usage()
+    iterator = _make_iterator(
+        sse_events=[
+            _sse_event({"type": "response.output_text.delta", "delta": "hello world"}),
+            _sse_event({"type": "response.completed", "response": {}}),
+        ],
+        logging_obj=_logging_obj_stub(),
+        config=_mock_config_with_completed_response(response),
+        request_data={"input": "count these input tokens please"},
+    )
+
+    with patch.object(litellm, "token_counter", side_effect=RuntimeError("tokenizer exploded")):
+        yielded: list = []
+        async for chunk in iterator:
+            yielded.append(chunk)
+
+    assert yielded
+    assert iterator.completed_response.response.usage is None