Anthropic's effort guide says to implement on low effort and verify on high, and that effort buys edge-case testing rather than a better approach
claude.dev blog: Using Claude Code: Spending your effort (corroborated by @trq212 on X)·high signal
Thariq Shihipar's 25 Sep claude.dev post finds that effort mainly controls how much verification and edge-case testing Claude does. On Terminal-Bench 3.0, Fable 5.1 used a median 73k tokens per attempt at low versus 222k at max, and higher effort lifted security tasks from 64% to 87% and hardware tasks from 34% to 75%. His workflow is to have Claude interview you against a spec, implement and iterate on low, then switch with /effort to high for verification and tests. Effort fixes missed edge cases, not a wrong approach, so when a task still fails at high, write a better spec instead of reaching for max.