Definition
A reasoning token is a provider-reported token used for intermediate generation by a reasoning model before or alongside its visible answer. It is an API accounting category, not a standard unit shared by every model provider.
OpenAI reports reasoning tokens within output-token details. Its documentation states that these tokens are not visible through the API, still occupy context-window space, are billed as output, and count toward max_output_tokens. Other providers may expose thinking text, summaries, budget controls, or no separate count at all. Read the contract for the exact model and endpoint.
Reasoning tokens are also not a faithful transcript of a model's causal process. A count tells an operator how much provider-defined intermediate generation occurred. It does not prove that the model checked the right facts, followed a sound argument, or used the tokens efficiently.
Operational significance
Hidden reasoning can make visible output length a poor estimate of cost or capacity. A low maximum-output setting can be exhausted before the answer begins, and a short answer can carry a large output bill. Track detailed usage, incomplete-response reasons, and total multi-call spend when setting budgets.
Distinguish it from nearby terms
- Output tokens are the broader generated-token category in APIs that include reasoning inside output usage.
- Chain-of-thought is a form of intermediate reasoning content. A reasoning-token count does not necessarily expose that content.
- Test-time compute includes more than token generation and is the broader resource concept.
- A reasoning-effort setting is an input control. Reasoning tokens report what the model used under provider-specific accounting.
Check your understanding
A response contains 300 visible tokens and 4,000 reported reasoning tokens. Explain which number matters for the user interface, which matters for billing, and why neither proves correctness.