Reasoning tokens
Last verified
Reasoning tokens are the intermediate work a reasoning model produces before its final answer. They are billed as output tokens and are not returned in the response body.
This is the most common cause of "why is my bill so much larger than the text I got back?"
Two things to watch:
- Effort settings multiply cost. Higher reasoning effort means more internal tokens.
- Usage data reveals them. The response's usage figures report the true output count, so compare that against the visible text if a bill looks wrong.
Use a reasoning model where reasoning helps. On extraction, classification and formatting, it usually adds cost without adding quality.