fix(openai): stop dropping cost when usage is a dict or cost is an int

[arthi-arumugam-git]

What

_parse_cost in langfuse/openai.py discards cost in two shapes that occur in practice, so cost_details comes back None and the invocation is recorded with no cost.

if hasattr(usage, "cost") and isinstance(getattr(usage, "cost"), float):

1. usage as a dict. hasattr({"cost": 0.002}, "cost") is False, so cost is dropped entirely. This is not hypothetical: _parse_usage right above it already branches on isinstance(usage, dict), and _update_langfuse_generation calls both with the same object:

if usage is not None:
    update["usage_details"] = _parse_usage(usage)
    update["cost_details"] = _parse_cost(usage)

So any dict-shaped usage that _parse_usage handles correctly loses its cost.

2. An integer cost. isinstance(0, float) is False in Python. OpenRouter reports "cost": 0 for free models and for fully cached responses, and any whole-number cost serialises as an int. Those get dropped too.

Measured against the current code, only one of five realistic shapes survives:

input	before	after
Usage(cost=0.0021)	{"total": 0.0021}	{"total": 0.0021}
Usage(cost=0)	None	{"total": 0.0}
Usage(cost=2)	None	{"total": 2.0}
{"cost": 0.0021}	None	{"total": 0.0021}
{"cost": 0}	None	{"total": 0.0}

Change

Read cost off either shape, and accept int as well as float. bool is excluded explicitly because it is a subclass of int, so a stray "cost": true would otherwise be recorded as a cost of 1.0.

Tests

tests/unit/test_parse_cost.py, 8 cases. Two of them fail on the current code and pass with the change:

FAILED tests/unit/test_parse_cost.py::test_dict_with_float_cost
FAILED tests/unit/test_parse_cost.py::test_integer_cost_is_not_dropped
2 failed, 6 passed

The other six cover None usage, an object without cost, a dict without cost, and the bool and string rejections, so the fix cannot widen the accepted set by accident.

ruff format --check is clean on both files. The pre-existing ruff check findings in openai.py are untouched.

Greptile Summary

This PR fixes OpenRouter cost extraction and expands regression coverage.

• Reads cost from both dictionary-shaped and object-shaped usage payloads.
• Accepts integer and floating-point costs while explicitly rejecting booleans.
• Normalizes accepted costs to floats and adds tests for supported and rejected values.

Confidence Score: 5/5

The PR appears safe to merge with no actionable defects identified.

The parser now handles the documented usage shapes while preserving rejection of unsupported values, and the downstream cost-details contract accepts the normalized floating-point result.

_{Reviews (1): Last reviewed commit: "fix(openai): stop dropping cost when usa..." | Re-trigger Greptile}

Context used:

• Knowledge Base — OpenAI Integration (langfuse/openai.py)

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.