Opus 5.5 feels worse? It's the medium effort default, not a nerf

Two independent trackers show no real drop since launch — the real causes are a lowered reasoning default, silent safety routing, and a thinking-content bug.

Nowline NOV 10 9:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • No, the trackers don't show a nerf

    Two continuous benchmarks that re-run frozen question sets — LiveNerf and BridgeBench — have stayed inside the normal 90–110% band since Opus 5.5's Sept 22 launch (BridgeBench read 94.2% on Oct 2 and 105.8% on Oct 16; LiveNerf's first window: +0.007 ± 0.018, "no change"). Each 10-day window can still miss a ~6-point slide, so keep your own frozen eval set instead of trusting vibes.

  • The real culprit: a lower effort default

    Opus 5.5 defaults to medium reasoning effort in Claude Code and the API, where Opus 5 defaulted to high and the launch benchmarks ran at max. Medium scores ~4.2 points lower while spending 26% fewer output tokens, and a legacy global "high" setting does not carry over to 5.5. Check with /effort status and raise it for hard tasks.

  • Silent safety routing can strand your session

    Prompts flagged as cybersecurity reroute to Opus 4.8 and biology to Opus 5 — and the session can stay on that fallback afterward, so later turns silently run on the older model instead of 5.5. Confirm what's actually serving you with /model or the modelUsage field in JSON output; claude --safe-mode isolates it.

  • Update Claude Code if you're below 2.1.293

    Builds 2.1.287 through 2.1.293 patched bugs that dropped thinking content and re-ran already-finished work after context compaction — a genuine quality hit that reads like a nerf. Upgrade past 2.1.293 and run /clear when you switch tasks in a long session.