GPT-oss-120b Reasoning Tokens Not Counted in Responses API Usage Statistics

Forum
1

I’m using the Groq Responses API with GPT-OSS models and noticed that reasoning tokens are not being counted in the usage statistics, even when reasoning is enabled and working correctly.

Environment:

Issue:
When making requests with reasoning enabled (reasoning: { effort: ‘high’ }), the API:

  1. :white_check_mark: Successfully generates reasoning content (visible in output array with type: ‘reasoning’)
  2. :white_check_mark: Counts total tokens in usage.input_tokens and usage.output_tokens
  3. :cross_mark: Always returns 0 for usage.input_tokens_details.reasoning_tokens and
    usage.output_tokens_details.reasoning_tokens

Example Request:
{
“model”: “openai/gpt-oss-120b”,
“input”: “What is 2 + 2? Think step by step.”,
“instructions”: “You are a helpful assistant.”,
“reasoning”: { “effort”: “high” },
“temperature”: 0.7
}

Actual Response (abbreviated):
{
“output”: [
{
“type”: “reasoning”,
“content”: [{“type”: “reasoning_text”, “text”: “Let me calculate…”}]
}
],
“usage”: {
“input_tokens”: 599,
“output_tokens”: 40,
“total_tokens”: 639,
“input_tokens_details”: {
“cached_tokens”: 0,
“reasoning_tokens”: 0 // ← Should be non-zero
},
“output_tokens_details”: {
“cached_tokens”: 0,
“reasoning_tokens”: 0 // ← Should be non-zero
}
}
}

Expected Behavior:
The reasoning_tokens fields should contain the actual count of tokens used for reasoning, separate from the
regular response tokens.

Questions:

  1. Is this a known limitation of the current beta implementation?
  2. Are there plans to implement proper reasoning token counting?
  3. Is there anything I need to configure differently to get reasoning tokens counted?

Thank you for your help!

2

Thanks for this report and the detailed information! This looks like it’s an issue on our end - I’m working with the team now to fix this.

I’ll post here again once it’s fixed!

3

Brilliant - thank you!