Reasoning Tokens
The hidden "thinking" tokens a reasoning model generates before its answer — billed at the output rate even though you never see them.
Reasoning tokens are the internal chain-of-thought a reasoning model produces while working through a problem before it writes the visible answer. Models like OpenAI’s o-series and Mistral’s Magistral spend these tokens to plan, check, and revise — and they are billed at the output rate even though they are not returned in the response.
This makes reasoning models cost more per task than their sticker output rate suggests, because a single answer can consume many multiples of the visible output in hidden reasoning. A model that looks cheap on paper can be expensive in practice on hard problems where it thinks for a long time, and cheaper on easy ones where it thinks briefly.
The upside is accuracy: on math, coding, and multi-step reasoning, spending tokens to think often beats a larger non-reasoning model. When budgeting, treat the output rate as a floor, not the true cost, and test on representative prompts to measure real token usage. Some providers expose controls to cap or tune reasoning effort — use them to keep spend predictable. This explorer tags reasoning-tier models so you can see them at a glance.
FAQ
Do I pay for reasoning tokens I never see?
Yes. The hidden reasoning a model generates is billed at the output rate, so a reasoning model’s real cost per answer can far exceed its visible output length.
Are reasoning models always more expensive?
Per task, often yes on hard problems, because they generate many hidden tokens. On simple prompts they think less and cost less. Test on your own prompts to gauge real usage.