Explicit high-effort reasoning flag on Claude Sonnet 5 adds $0.01031 per call with no measurable accuracy gain on math reasoning tasks.
Summary
If you're already paying for extended thinking on production math/logic workloads, the explicit high-effort contract may be a cost sink. Understanding when effort flags actually move accuracy helps right-size inference spend.
Why it matters
If you're already paying for extended thinking on production math/logic workloads, the explicit high-effort contract may be a cost sink. Understanding when effort flags actually move accuracy helps right-size inference spend.
Implementation verdict
This replaces assumption-based reasoning spend tuning with empirical cost-benefit data. Requires pre-registering your own contrasts on representative tasks before committing to high-effort billing. Not ready as a general rule—task and model dependent, findings bounded to AIME 2026 and Sonnet 5 at test date.
Sources
Dev Signal
Get briefs like this in your inbox — free, every weekday.
100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.