Reasoning effort, not temperature, now tunes AI prompts

Reasoning models swapped their main tuning knob this year, from temperature to effort, per IBM.

Get the level wrong and it costs you twice. Multi-step problems run thin on low effort, simple lookups burn tokens they never needed on high.

Claude Code already ships four effort levels, low, medium, high, and max, plus adaptive thinking that adjusts on its own. High effort fits multi-step logic, the kind of work that needs the model to plan several moves and backtrack when it's wrong. Low effort fits retrieval and formatting, pulling a value out of text or reshaping it.

Effort tuning is becoming the real skill in prompting reasoning models, ahead of clever prompt wording. Watch whether adaptive thinking makes the manual dial optional, or whether picking effort by hand stays how you get consistent results.

This is the year reasoning models made effort the main dial, per IBM. Claude Code's four-level setting is already built for it.