Users Claim Anthropic Is Quietly Throttling Claude Code's 'Effort' in an A/B Test

A widely-shared post alleges that Anthropic is running a server-side experiment that compresses the ‘effort’ scale for Fable 5 sessions in recent builds of Claude Code (2.1.236 and later). The claim is that enrolled users who select ‘high’ effort actually get behavior closer to ‘low’ — fewer reasoning steps, more consolidated tool calls, terser output — while older client versions and Opus-tier sessions are left untouched. The poster stresses the change is delivered on Anthropic’s side, not through the app binary, and that only a fraction of users are bucketed into the test, which would explain why some see degraded results and others don’t.

The detail that gives the accusation teeth is that ‘effort’ is a real, load-bearing control in Claude’s stack. It governs how deeply the model reasons and how much it spends per turn, and higher tiers are the default for agentic coding work. Silently narrowing that scale would let a request labeled ‘high’ quietly execute like a cheaper setting — cutting inference cost without a visible switch — which is exactly the kind of invisible quality regression that erodes trust in a paid developer tool. If accurate, the story is less about a bug than about undisclosed experimentation on output quality.

That said, this rests on a single unverified social-media report with no confirmation from Anthropic and no reproducible measurements attached. One detail undercuts its precision: the post says ‘Opus 5’ is exempt, but there is no Opus 5 in Anthropic’s current lineup — the top Opus model is Opus 4.8, alongside Fable 5. Readers should treat the specific version numbers and model names as approximate until someone publishes side-by-side telemetry or Anthropic responds.