Thinking Tokens
Tokens generated by an AI model during an extended reasoning process that are used internally for step-by-step problem-solving but are not shown to the end user. Also called 'reasoning tokens' or 'thinking budget,' these tokens represent the model's intermediate reasoning—similar to scratch work on a math problem. Models like Claude (extended thinking) and OpenAI's o-series use thinking tokens to improve accuracy on complex tasks. Thinking tokens count toward the model's output token usage and billing, but give the model space to reason through multi-step problems, check its work, and consider alternatives before producing a final answer.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A legal agent asked to analyze a complex contract clause uses 2,000 thinking tokens to work through the implications: identifying relevant precedents, considering edge cases, evaluating risk from multiple angles, and checking its reasoning before producing a concise 200-token analysis for the lawyer. The thinking tokens aren't shown to the user but dramatically improve the quality and accuracy of the final output.
Frequently asked questions
- Do thinking tokens cost extra?
- Thinking tokens are billed as output tokens since the model generates them. A task that produces 200 visible tokens might use 2,000 thinking tokens behind the scenes—so the total billed output is 2,200 tokens. This makes reasoning-heavy tasks more expensive per visible token, but the accuracy improvement often justifies the cost. Some providers offer thinking token budgets to cap this cost.