Opus 5 comes with five effort settings: low, medium, high, xhigh, max. Most people seem to be reaching straight for max, and at least on coding work that looks like the wrong move.
On FrontierCode, scores fall above the high setting. The stated reason is that the model starts making unnecessary refactors and edits outside the scope it was given. Anthropic's own migration guide in the system card warns about diminishing returns and overthinking on simpler tasks, so this is not some outside critic's claim.
Two other numbers point the same way:
- On the closed-book AA-Omniscience benchmark, Opus 5 is about 11% more accurate than Opus 4.8, but its hallucination rate runs about 6% higher. More reasoning, more room to be confidently wrong.
- CodeRabbit ran it at xhigh against their production baseline for code review. Precision on actionable comments went up, 39.3% vs 35.2%. But it caught fewer of the benchmark's known issues, 55.2% vs 61.1%, and generated roughly four times as many nitpicks.
The flip side is worth knowing too, because it cuts the other way. On Zapier's AutomationBench, Opus 5 at its lowest effort setting still passes more tasks than any other model. So for a lot of workloads the cheap end of the dial is already enough, and the expensive end is not just wasted spend, it can be actively worse output.
So, the setting where Opus 5 stops improving is probably specific to your codebase, and nobody has published a map of it. Worth finding your own ceiling before you default everything to max.
One unrelated thing I have not seen discussed much: when a safety classifier flags a request in Claude.ai, Claude Code or Cowork, it silently falls back to Opus 4.8 by default. That is also how Anthropic's own Frontier-Bench run was configured, per the footnote on their chart. Nobody has published what fraction of requests that affects.
Has anyone found the effort level where it turns over on a real repo? Curious whether the drop-off point moves with codebase size or with how much context you hand it.
[link] [comments]