← Back to glossary
+Suggest a term
Concept·Patterns & Practices·Added 1 month ago

Token Budget

Also known as: thinking budget, reasoning budget, compute budget

A limit you set on how many tokens a model can spend on internal reasoning before producing an answer. Higher budgets mean more thorough reasoning at the cost of latency and money. Lower budgets are faster and cheaper. Most useful when using extended thinking or reasoning models.

Reasoning models and extended thinking modes don't just generate an answer; they first generate a chain of internal thought. All those thinking tokens cost money and time. A token budget lets you cap that spend. Set it high for complex problems where accuracy matters; set it low (or to zero) for simple tasks where thinking is wasteful.

Anthropic exposes this in the Claude API as a thinking parameter with a budget_tokens field. OpenAI has analogous reasoning_effort settings (low, medium, high) on o-series models. Google offers similar controls on Gemini's Deep Think mode. The interface differs but the concept is the same: you're buying more or less deliberation time.

Token budgets are a new dimension of production cost management that didn't exist before reasoning models. Builders optimizing agent pipelines now need to think not just about input/output token counts but also about thinking token costs per step. A multi-step agent that uses high thinking budgets on every step can generate costs that dwarf simple API call patterns.

This definition is AI-generated and refreshed weekly. It may contain inaccuracies. Use your own judgment, especially for production decisions.
Related terms
Extended ThinkingInference-time computeTokenContext compaction