Effort setting
An API parameter that controls how much a model reasons: low effort answers faster and cheaper, high effort is slower and more thorough.
Why it matters. Tune it per route: low for chat and classification, high for agents and hard problems.
How it works
An effort setting is an API parameter that tells a reasoning model roughly how much work to put into a request, usually as levels such as low, medium and high. Lower effort means fewer reasoning tokens, faster replies and lower cost; higher effort means more thorough work on hard problems.
Makers name and implement it differently: some use named levels, others a numeric thinking budget, and defaults differ by model. Some models also let you switch reasoning off; others always reason. Effort is guidance rather than an exact count, so token use still varies per request.
Set effort per route rather than globally, and evaluate quality at each level on your own tasks. The highest level can cost much more for little gain on easy work.
Worked example
Illustrative: a coding route uses about 8,000 reasoning tokens per request at high effort and 2,000 at low. At $10 per 1M output tokens and 10,000 requests a day, that is $800 versus $200 a day in reasoning alone.
Common mistakes
- Leaving the default effort on every route instead of tuning per task.
- Treating effort as a hard token cap; set max output tokens as well.
- Assuming effort levels mean the same thing across makers or model versions.
See it on a real card: Specifications on Claude Opus 5.5.