On reasoning models, tune reasoning_effort instead of temperature

IBM·medium signalXBlueskyLinkedInCopy link

For 2026 reasoning models, the primary control lever has shifted from temperature to a reasoning_effort parameter (low/medium/high): raising it spends more tokens on hidden chain-of-thought and sharply improves logic accuracy on hard problems, while lowering it saves cost on easy ones. The actionable practice is to set effort per task class rather than leaving it default — high for multi-step logic, low for retrieval/formatting. Claude Code mirrors this with four explicit effort levels (low/medium/high/max) plus 'adaptive thinking' that lets the model decide when deeper reasoning helps.