Reasoning tokens: what a thinking model spends before it answers
Reasoning tokens (thinking tokens) are the tokens a reasoning model generates in a separate thinking phase before its visible answer. They are produced like any other token, so they cost generation time and fill the context window, even when the client hides them. Many models and servers let you cap them with a thinking budget or switch thinking off.
At a glance
What they are
Generated text in a thinking phase, before the answer
What they cost
Generation time and context, like visible tokens
Control
A thinking budget, or thinking on or off per request
Failure mode
Overthinking: the budget runs out before any answer
How is this different from chain of thought?
Chain of thought is a prompting technique: you ask the model to show its steps in
the answer. Reasoning tokens come from models trained to think by themselves in
a separate phase that the server can parse out and the client can hide. The
effect is similar, more deliberate answers, but the tokens are spent whether you
see them or not, and they count against the same budget.
How do I keep the cost in check?
Measure how many thinking tokens your tasks really use, then cap them. Turn
thinking off for short, mechanical requests and on for hard ones, and set a
timeout or budget so a model stuck in a loop fails fast instead of spending
minutes. A model whose median thinking is ten times its answer is expensive even
when its answers are good.