Claude Effort Levels Benchmarked: Low vs Max on Real Tasks

A developer tested Claude's five effort settings on classification, code generation, and multi-step audit tasks. For simple classification, low effort matched max quality but used 8x fewer tokens. For complex audits, xhigh used fewer total tokens than medium by avoiding dead ends.