Report: Anthropic appears to A/B test effort levels in Claude Code
Builders report Opus 5 burning tokens and time on trivial tasks, with no word from Anthropic. Plus: Cursor's agents go always-on and Kimi splits its plans.
Copy markdown
Ask Claude what effort it's running on
Developers on Hacker News (83 points in ~3 hours) report Claude Code's thoroughness swings between sessions, and that Opus 5 will name an "effort level" when asked directly — the tell of a live A/B test. Worth checking which bucket you're in before you blame your own prompt.
The 43-minute config edit
One user reports Opus 5 spent 43 minutes spinning up sandboxes and test suites for a config change Opus 4.6 finished in under two minutes with identical output. Several say they've cancelled $200+ Max plans or moved to Codex and GLM-5.3 over the token bloat.
No confirmation — and a familiar pattern
Anthropic hasn't acknowledged an effort experiment. It echoes the spring degradation saga, which Anthropic eventually traced to harness and config bugs — not a weaker model — in a public postmortem. Until there's official word, treat this as reportedly.
Elsewhere: Cursor's cloud agents go always-on
Cursor's latest update gives cloud agents Subscriptions that wake them on PR, Slack and scheduled events, runs each subagent in its own isolated VM, and adds a /goal command for objectives an agent chases until done. A real step toward set-and-forget background agents.
Elsewhere: Kimi is about to split its plans
As of Aug 20, kimi.com shows a banner that Kimi and Kimi Code benefits will separate into distinct memberships, existing subscribers unaffected. If you lean on Kimi K3 for coding ($3/$15 per M tokens on the API), expect to pick a tier soon.