Anthropic appears to be A/B testing reduced effort levels in Claude Code
Points and comments are a snapshot, not live.
Anthropic appears to be A/B testing reduced effort levels in Claude Code, making "high" feel like "low".
A user reports that Anthropic is running a server-side A/B test on Claude Code version 2.1.236+, shrinking the effort scale so that "high" effort now corresponds to a value of 10 out of 100, the same number "low" used to be. Older versions and Opus 5 are not affected. The changelog does not mention the change. The user spent an afternoon believing their own code was broken before discovering the experiment.
What commenters are saying
Many commenters suspect broader optimization tactics beyond effort scaling, including dynamic model routing and usage-limit rubberbanding. Several report that API-based models feel unaffected compared to chat interfaces. A user claims Opus 5 spent 43 minutes on a simple file modification that took Opus 4.6 under 2 minutes. Others note the model's verbosity has increased significantly, and some have cancelled subscriptions due to perceived regression. One camp argues companies have financial incentives to burn tokens; another counters that making models faster adds more user value.