Reasoning tokens

Tokens a model generates while working through a problem before answering. You are billed for them as output, even when they aren't shown.

Why it matters. They improve hard tasks but add cost and latency; budget for them.

How it works

Reasoning tokens are tokens a model generates while working through a problem before writing its final answer. They are billed at the output rate and count toward output limits, even when the API returns only a summary of the reasoning or none at all.

Makers expose reasoning differently: some return the full thinking text, some a summary, some only a count in the usage data. The amount varies with task difficulty and with the effort or budget setting.

Reasoning helps most on math, coding, planning and multi-step agent work, and often adds little for simple lookups or rewriting. Compare tokens and quality per task with reasoning on and off before deciding.

Worked example

Illustrative: a request returns a 400-token answer but the usage data shows 3,600 reasoning tokens. At $10 per 1M output tokens you pay for 4,000 tokens, $0.04, ten times the $0.004 the visible answer suggests.

Common mistakes

Estimating cost from the visible reply instead of the usage data.
Setting max output tokens so low that reasoning uses it up before the answer is written.
Leaving heavy reasoning on for simple, high-volume routes such as classification.

See it on a real card: Specifications on Claude Opus 5.5.