No access to reasoning_tokens via Gemini API's OpenAI compatible layer

Gemini API
1

Issue

Gemini API’s OpenAI compatible layer does not expose reasoning_tokens. This is making it difficult for applications to control reasoning.

To replicate, make any request to a reasoning model (e.g. Gemini 2.5 Flash) via the OpenAI compatible layer.

The usage field does not directly mention reasoning tokens

    "usage": {
        "completion_tokens": 18,
        "prompt_tokens": 15,
        "total_tokens": 175
    }

Compare this to a response from the same model in Vertex AI:

    "usage": {
        "completion_tokens": 21,
        "completion_tokens_details": {
            "reasoning_tokens": 78
        },
        "extra_properties": {
            "google": {
                "traffic_type": "ON_DEMAND"
            }
        },
        "prompt_tokens": 14,
        "total_tokens": 113
    }

Suggested solution

Match the convention used by Vertex AI