DictionaryToken Economics
Thinking Tokens
The hidden scratch work the AI does before answering, which you pay for like regular output.
Definition
Tokens generated by a model's internal reasoning process before producing a visible response. Thinking tokens consume context budget and are billed as output, but can dramatically improve quality on complex tasks. On current Claude models the amount of thinking is controlled with adaptive thinking and effort rather than a budget_tokens cap.
Example
A reply shows 300 visible words, but usage reports 6,000 output tokens because most went to thinking, billed at the output rate.