Cost
Reasoning tokens
Reasoning tokens are the internal thinking a reasoning model generates before its answer. They are billed as output and never returned to you.
Also written: thinking tokens
At high reasoning effort the hidden portion can be several times the visible answer, which makes every budget built on answer length wrong in the same direction.
They also occupy the context window. On a long conversation with a reasoning model, thinking takes space you assumed was available for history or retrieved content.
In practice
You are billed for tokens you never see. A reasoning model asked a hard question can spend several thousand tokens thinking before it writes a 200-token answer, and the invisible tokens are charged at the output rate. Budgeting from visible output length on a reasoning model can understate the bill by an order of magnitude.
Common questions
Am I charged for reasoning tokens?
Yes, at the output rate, even though you never see them. On hard questions the invisible reasoning can outnumber the visible answer several times over, which is why budgeting from answer length understates the bill.
Can I control how many reasoning tokens a model uses?
On models that expose a reasoning effort setting, yes — lower it for tasks that do not need multi-step deduction. On tasks that are genuinely simple, a non-reasoning model is cheaper than a reasoning model set to low.