Claude Code quietly turned 'high' effort into what 'low' used to be
Anthropic's flagship coding tool appears to be running a silent experiment on the people who pay the most attention to it — and the only reason anyone knows is that one developer refused to believe his own setup was broken.
Fable 5 sessions in Claude Code have been reading the "high" effort setting as roughly a 10 out of 100 since version 2.1.237 — numerically the same spot "low" used to occupy, according to developer @argofowl, who traced the behavior after an afternoon of debugging convinced him something had changed. The remapping is applied server-side rather than in the app, so users enrolled in the test can't escape it by rolling back; older client versions and Opus 5 sessions appear untouched. Anthropic has not acknowledged the experiment, and the changelog says nothing about any change to how effort levels are interpreted.
Effort levels are not a cosmetic toggle. Per Anthropic's own documentation, effort governs how much work the model does on a request — how many files it reads, how many tools it uses, and how many steps it takes before checking back in. Quietly remapping it means paying users who explicitly selected maximum thoroughness got less of it, with no way to distinguish the drop from model drift or their own mistakes. That last part is the real sting: the developer's first assumption was that his code was broken, not the tool.
We covered Claude Code's drift toward invisible defaults earlier this month — Claude Code makes auto mode the default for paid plans — and this pushes further in the same direction. Silent A/B tests are routine on consumer web apps; they are contested territory for a tool wired into people's production workflows. When the product can rewrite its own behavior between two runs of the same prompt, reproducibility — the main thing a coding agent sells — becomes conditional on which experiment bucket you happen to land in.
There is a fair case for the test itself: high effort burns real compute, and plenty of prompts don't need it. But the failure here isn't experimentation — it's silence. One line in the release notes would have cost Anthropic nothing and spared users the week-long suspicion that their own projects were the problem.
What to watch: whether Anthropic confirms the experiment — and more importantly, whether it starts surfacing effective-effort telemetry so developers can verify what their settings actually do. Trust repairs faster with receipts than apologies.
Has your agent felt off lately — and would a silent downgrade like this push you to switch tools? Tell us in the comments.
