Home › Blog › Reasoning Tokens Cost Too Much: How to Control Them
OutputReasoning Tokens Cost Too Much: How to Control Them
Reasoning models think before they answer, and that thinking is billed. It can be most of a request's cost without appearing in the reply.
Thinking is billed as output tokens, the most expensive kind. Set effort or thinking budgets per task instead of leaving every call on the maximum.
How it's billed
Google's pricing page lists Gemini output prices as including thinking tokens. Reasoning on other providers is also billed at output rates. Output already costs several times input, so long reasoning on simple tasks is expensive.
Controls by provider
- OpenAI: reasoning effort settings on reasoning models.
- Anthropic: an effort setting on current Claude models; lower effort uses fewer thinking tokens.
- Gemini: thinking budgets on models that support them.
A practical approach
- Group requests by task type.
- Lower effort on simple tasks and compare pass rates on your evals.
- Keep high effort only where it measurably improves results.
- Cap total output with a sensible limit.
Thinking is often the hidden line item in an agent bill. See how to cap agent costs.
Frequently asked questions
Are thinking tokens billed?
Yes. Reasoning is billed at output rates; Google's pricing page says Gemini output prices include thinking tokens.
How do I reduce reasoning costs?
Lower the effort or thinking budget for simple tasks and test the impact on your evaluation set.
Sources: Google's Gemini API pricing page. Checked October 30, 2026.
Want us to find your 20%?
Book a free 15-minute API cost teardown. Bring last month's provider usage and we'll show you your biggest waste and the first fix we'd ship.
Get a Free 15-Minute API Cost Teardown