Reasoning models and effort settings

Some models think before they answer. They first write internal reasoning tokens, working through the problem, and only then write the final reply. Many let you set an effort level or a thinking budget, which caps how much they reason.

More reasoning helps on problems with many steps: maths, planning, tricky code and careful extraction from messy documents. But those reasoning tokens are billed, usually at the output price, and they add seconds of waiting. For simple classification or short rewrites, a fast model with no extra reasoning is often just as accurate and far cheaper.

Treat effort as one more setting to measure with your evals, not a quality dial to turn up by default.