fix(responses): estimate failed-stream usage without fetching image inputs

litellm/responses/streaming_iterator.py

@@ -498,6 +498,7 @@ class BaseResponsesAPIStreamingIterator:
                                     self.request_data.get("input"),
                                     self.request_data,
                                     self._generated_content + self._generated_tool_arguments,
+                                    use_default_image_token_count=False,
                                 )
                                 if _estimate_wanted
                                 else None
@@ -888,7 +889,11 @@ class BaseResponsesAPIStreamingIterator:
         if not generated_text or self.logging_obj.model_call_details.get("combined_usage_object") is not None:
             return
         estimate: Final = _estimate_usage_safely(
-            self.model or "", self.request_data.get("input"), self.request_data, generated_text
+            self.model or "",
+            self.request_data.get("input"),
+            self.request_data,
+            generated_text,
+            use_default_image_token_count=True,
         )
         if estimate is None:
             return
@@ -1558,13 +1563,15 @@ def _estimate_usage_from_text(
     request_input: object,
     responses_api_request: Mapping[str, object],
     generated_text: str,
+    *,
+    use_default_image_token_count: bool,
 ) -> ResponseAPIUsage:
     messages: Final = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(  # pyright: ignore[reportUnknownMemberType]  # the transformer's signature is partially untyped
         input=request_input,  # pyright: ignore[reportArgumentType]  # the raw Responses API input is a str or ResponseInputParam list, matching the helper's declared union
         responses_api_request=dict(responses_api_request),
     )
     input_tokens: Final = litellm.token_counter(  # pyright: ignore[reportUnknownMemberType]  # token_counter's public signature is untyped
-        model=model, messages=messages
+        model=model, messages=messages, use_default_image_token_count=use_default_image_token_count
     )
     output_tokens: Final = litellm.token_counter(  # pyright: ignore[reportUnknownMemberType]  # token_counter's public signature is untyped
         model=model, text=generated_text, count_response_tokens=True
@@ -1581,6 +1588,8 @@ def _estimate_usage_safely(
     request_input: object,
     responses_api_request: Mapping[str, object],
     generated_text: str,
+    *,
+    use_default_image_token_count: bool,
 ) -> ResponseAPIUsage | None:
     try:
         return _estimate_usage_from_text(
@@ -1588,6 +1597,7 @@ def _estimate_usage_safely(
             request_input=request_input,
             responses_api_request=responses_api_request,
             generated_text=generated_text,
+            use_default_image_token_count=use_default_image_token_count,
         )
     except Exception as e:
         verbose_logger.debug("Could not estimate usage from stream text, billing $0: %s", e)

tests/unit/llms/anthropic/pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py

@@ -658,10 +658,6 @@ def _collect_with_logging(stream, logging_obj: LitellmLogging) -> list:
 
 
 class TestUpstreamFailureReachesTheProxyFailureHook:
-    """The ``error`` event alone leaves the proxy unaware the stream failed, so the wrapper also hands
-    the failure to the hook the proxy arms on the logging object, the one that writes the failure
-    spend row and releases the budget reservation."""
-
     def test_raised_upstream_failure_reaches_the_hook_unwrapped_before_the_error_event(self):
         rate_limit = litellm.RateLimitError(message="You have no credits remaining.", llm_provider="openai", model="m")
         wrapped = MidStreamFallbackError(

tests/unit/responses/test_responses_streaming_iterator.py

@@ -183,11 +183,24 @@ class _ChunkRecorder:
             self.chunks = (*self.chunks, chunk)
 
 
+_SEA_DELTAS = ("The sea ", "is wide ", "and deep.")
+_SEA_PROMPT = "Write a 300 word story about the sea."
+_UNREACHABLE_IMAGE_INPUT = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "input_text", "text": "Describe this picture."},
+            {"type": "input_image", "image_url": "http://images.invalid/photo.png", "detail": "high"},
+        ],
+    }
+]
+
+
 class _UpstreamThatTimesOutAfterThreeDeltas:
     headers: dict = {}
 
     async def aiter_bytes(self):
-        for sequence_number, delta in enumerate(("The sea ", "is wide ", "and deep.")):
+        for sequence_number, delta in enumerate(_SEA_DELTAS):
             event = {
                 "type": "response.output_text.delta",
                 "item_id": "msg_1",
@@ -200,8 +213,7 @@ class _UpstreamThatTimesOutAfterThreeDeltas:
         raise httpx.ReadTimeout("Timeout on reading data from socket")
 
 
-@pytest.mark.asyncio
-async def test_mid_stream_read_timeout_logs_failure_with_partial_usage_and_no_success(monkeypatch):
+async def _failure_payload_after_mid_stream_timeout(monkeypatch, request_input) -> tuple[FailureRecorder, dict]:
     recorder = FailureRecorder()
     monkeypatch.setattr(litellm, "failure_callback", [recorder])
     monkeypatch.setattr(litellm, "_async_failure_callback", [recorder])
@@ -219,17 +231,50 @@ async def test_mid_stream_read_timeout_logs_failure_with_partial_usage_and_no_su
         responses_api_provider_config=OpenAIResponsesAPIConfig(),
         logging_obj=logging_obj,
         custom_llm_provider="openai",
+        request_data={"model": "gpt-5.4-nano", "input": request_input, "stream": True},
     )
     received = _ChunkRecorder()
 
     with pytest.raises(httpx.ReadTimeout, match="Timeout on reading data from socket"):
         await received.drain(iterator)
-    assert [chunk.delta for chunk in received.chunks] == ["The sea ", "is wide ", "and deep."]
+    assert [chunk.delta for chunk in received.chunks] == list(_SEA_DELTAS)
 
     await iterator._await_pending_logging()
     (payload,) = recorder.failure_payloads
+    return recorder, payload
+
+
+@pytest.mark.asyncio
+async def test_mid_stream_read_timeout_logs_failure_with_partial_usage_and_no_success(monkeypatch):
+    recorder, payload = await _failure_payload_after_mid_stream_timeout(monkeypatch, _SEA_PROMPT)
+
     assert payload["status"] == "failure"
-    assert payload["prompt_tokens"] > 0
-    assert payload["completion_tokens"] > 0
+    assert payload["prompt_tokens"] == litellm.token_counter(
+        model="gpt-5.4-nano", messages=[{"role": "user", "content": _SEA_PROMPT}]
+    )
+    assert payload["completion_tokens"] == litellm.token_counter(
+        model="gpt-5.4-nano", text="".join(_SEA_DELTAS), count_response_tokens=True
+    )
     assert payload["response_cost"] > 0
     assert recorder.success_payloads == ()
+
+
+@pytest.mark.asyncio
+async def test_mid_stream_failure_usage_estimate_never_fetches_image_dimensions(monkeypatch):
+    _, payload = await _failure_payload_after_mid_stream_timeout(monkeypatch, _UNREACHABLE_IMAGE_INPUT)
+
+    assert payload["status"] == "failure"
+    assert payload["prompt_tokens"] == litellm.token_counter(
+        model="gpt-5.4-nano",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Describe this picture."},
+                    {"type": "image_url", "image_url": {"url": "http://images.invalid/photo.png", "detail": "high"}},
+                ],
+            }
+        ],
+        use_default_image_token_count=True,
+    )
+    assert payload["completion_tokens"] > 0

tests/unit/responses/test_streaming_iterator.py

@@ -1141,7 +1141,9 @@ async def test_completed_event_survives_a_failing_usage_estimate():
     falls back to usage None."""
     malformed_input: Final = [{"type": "message", "role": "user", "content": 42}]
     with pytest.raises(ValueError, match="Invalid content type"):
-        _estimate_usage_from_text("gpt-4o-mini", malformed_input, {"input": malformed_input}, "hello world")
+        _estimate_usage_from_text(
+            "gpt-4o-mini", malformed_input, {"input": malformed_input}, "hello world", use_default_image_token_count=False
+        )
 
     response = _responses_api_response_without_usage()
     iterator = _make_iterator(