Bug: NoneType error in _get_attributes_from_response_usage for non-reasoning models

[danialm]

Description

_get_attributes_from_response_usage in openinference-instrumentation-openai crashes with AttributeError: 'NoneType' object has no attribute 'reasoning_tokens' when the OpenAI API returns output_tokens_details: None in usage data. This happens with non-reasoning models (e.g. GPT-4o) where output_tokens_details is not populated.

The same issue exists for input_tokens_details.cached_tokens.

Stack trace

Traceback (most recent call last):
  File "openinference/instrumentation/openai/_attributes/_responses_api.py", line 31, in wrapper
    yield from wrapped(*args, **kwargs)
  File "openinference/instrumentation/openai/_attributes/_responses_api.py", line 720, in _get_attributes_from_response_usage
    obj.output_tokens_details.reasoning_tokens,
AttributeError: 'NoneType' object has no attribute 'reasoning_tokens'

Affected versions

Confirmed in 0.1.41 and 0.1.44 (latest). The code is unchanged between versions.

Impact

Low — @stop_on_exception catches the error and logs it, so requests succeed. But it generates noisy ERROR logs and loses the reasoning/cache token count telemetry attributes.

Suggested fix

Add null guards before accessing nested attributes:

@classmethod
@stop_on_exception
def _get_attributes_from_response_usage(
    cls,
    obj: responses.response_usage.ResponseUsage,
) -> Iterator[Tuple[str, AttributeValue]]:
    yield SpanAttributes.LLM_TOKEN_COUNT_TOTAL, obj.total_tokens
    yield SpanAttributes.LLM_TOKEN_COUNT_PROMPT, obj.input_tokens
    yield SpanAttributes.LLM_TOKEN_COUNT_COMPLETION, obj.output_tokens
    if obj.output_tokens_details:
        yield (
            SpanAttributes.LLM_TOKEN_COUNT_COMPLETION_DETAILS_REASONING,
            obj.output_tokens_details.reasoning_tokens,
        )
    if obj.input_tokens_details:
        yield (
            SpanAttributes.LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_READ,
            obj.input_tokens_details.cached_tokens,
        )

[dosubot]

Hi @danialm! I'm Dosu and I’m helping the openinference team.

This was fixed in PR #2496, which was merged on December 3, 2025. The fix adds the exact null guards you suggested — checking if obj.output_tokens_details and if obj.input_tokens_details before accessing the nested attributes [1].

However, since you're seeing this on v0.1.44 (described as the latest), it's possible this fix hasn't been included in a release yet. You may want to check if there's a newer version available, or install from the main branch to pick up the fix.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}