Claude Code Users Flag Hidden Effort Cuts

Anthropic’s effort controls are becoming an opaque product lever: users can’t meaningfully manage quality or cost if “high” changes underneath them without notice.

• –Community testing reportedly found high effort falling from 40 to 10 reasoning units across recent builds
• –The server-side mapping means identical client requests can produce materially different model behavior
• –Anthropic’s documentation acknowledges that effort levels are calibrated per model, but does not explain silent build-to-build changes
• –Lower reasoning budgets may improve latency and usage economics while hurting long-horizon coding and multi-file refactors
• –Developers should benchmark fixed tasks across Claude Code versions instead of trusting the displayed effort level