What is Reasoning Effort?
Also called thinking budget, reasoning mode.
Reasoning effort is a setting on reasoning-capable models that controls how much internal deliberation a model performs before answering. Higher settings let it generate more hidden reasoning tokens, which improves accuracy on hard problems while increasing latency and cost. Providers expose it either as named levels such as low, medium, and high, or as a numeric token budget.
Underneath, the setting governs how many reasoning tokens the model may produce before committing to a response. Those tokens are generated and billed like other output tokens even when they are never shown to the user. Raising the budget on a difficult problem often turns a wrong answer into a right one; raising it on a trivial one changes nothing but the bill.
The right level is task-dependent and worth measuring rather than assuming. Classification, extraction, formatting, and simple factual lookups rarely benefit at all. Multi-step arithmetic, debugging, planning, and problems with several interacting constraints benefit substantially. Applications handling mixed traffic frequently set the level per request rather than choosing one global value for everything.
Reasoning tokens interact with the context window, because they occupy part of the same budget. A high effort setting combined with a very long input can exhaust the window before the final answer is produced, resulting in a truncated or empty response. Provider limits on visible output are typically counted separately from the reasoning allowance itself.
The feature also changes what other settings do. Several providers restrict or ignore temperature and top-p on reasoning models, since sampling is managed internally by the model. Hidden reasoning is generally not returned verbatim, so it cannot serve as an audit trail, and any summary of it is a reconstruction rather than a faithful transcript.
Key points
- Controls how much hidden deliberation happens before the answer.
- Reasoning tokens are billed even when they are never displayed.
- Helps on multi-step problems; wasted on simple extraction or lookup.
- Competes with input for the context window and can truncate answers.
- Often disables or overrides temperature and top-p settings.
In practice
A model is asked to reconcile two spreadsheets with mismatched date formats and one duplicated row. At the lowest effort it returns a fast, confident, wrong total. At a higher setting it spends several thousand hidden tokens comparing rows, notices the duplicate, and returns the correct figure with the discrepancy flagged. The answer took longer and consumed noticeably more tokens.