Claude Code users report an undisclosed reasoning-effort throttle

Some Claude Code users are reporting a possible behind-the-scenes experiment that changes how much reasoning effort the model applies to a task. The claim is that versions 2.1.236 and later may silently place some sessions on a smaller effort scale, without showing the user that anything has changed.

One reported example is striking, but still needs treating with caution. A user said Opus 5 took 43 minutes on a task that Sonnet 4.6 completed in under two minutes. That could point to a model-routing or reasoning-setting issue, but it is only a single-thread report so far. Anthropic hasn’t commented, and there’s no changelog entry confirming an experiment. The timing claim is specific enough to test, but it isn’t proof on its own.

If your team uses Claude Code for development, don’t assume model names alone tell you what performance you’ll get. Track task duration, cost, completion quality, and retries by model and version. A sudden slowdown might be a prompt problem, a codebase problem, or a product-level change you weren’t told about. You need enough operational data to separate those possibilities before changing tools or workflows.

This is the kind of thing we build into an AI command centre, giving teams a practical view of model performance rather than relying on vendor labels or anecdotal reports.