A reasoning token is part of the text a reasoning model generates to think through a problem before it answers or calls a tool. Reasoning tokens count as output and fill the context window, but many hosted model APIs do not return them in full: they return a summary or nothing, which limits what a trace can contain.
Why it matters when licensing company data
Provider documentation is explicit. One major API says reasoning tokens are not visible via the API and offers a summary instead; another says the text in a thinking block is a summary of the model's reasoning, and hides even that by default on its newest models; a third returns thought summaries only when they are enabled. Open-weight models run on your own infrastructure can return the full reasoning.
Reasoning matters because it shows why an agent chose each step. For anyone describing a set of traces, the honest field is reasoning visibility: full, summary or none for every run. The OpenTelemetry GenAI conventions also say the output token count should include reasoning tokens, so token counts and visible text can differ.
See licensing agent traces from production and how to export agent traces.
Example
A team running two agents compares their traces. The agent on a hosted reasoning model shows high output token counts but only short summaries between steps. The agent on an open-weight model run in-house shows the full reasoning text. When the team builds a data card for each set, it marks the first as reasoning summarised and the second as reasoning full, and it keeps the token counts as reported rather than recounting them from the visible text.
Source
OpenAI API docs: reasoning models. Definition reviewed November 9, 2026 by Origin Data Partners.