ClaudeFolio
General

Has Opus 5.5 been nerfed? How to tell a real downgrade from a changed setting

Edward Kwun··4 min read
Has Opus 5.5 been nerfed? How to tell a real downgrade from a changed setting

See more of our writing in your Google results.

Key points

  • Anthropic says it never lowers quality for demand or load
  • Three real infrastructure bugs degraded Claude in 2025
  • Opus 5.5 defaults to Medium effort, below Opus 5
  • A flagged request can leave your session on Opus 4.8
  • Test with a pinned model, fixed effort, and known tasks

If you check social media or the Claude subreddits, you'll always see how people say a particular model has been downgraded or nerfed. However, Anthropic has said publicly that "we never reduce model quality due to demand, time of day, or server load," and most of the times Claude Opus 5.5 feels worse, something in your setup changed: the effort level, the model your session is actually running, or how full the conversation is. Real regressions do happen, though, so it's worth knowing how to tell the two apart.

That quote comes from Anthropic's postmortem of three bugs that degraded Claude in August and September 2025. One misrouted some Claude Sonnet 4 requests to servers set up for a different context length, starting at about 0.8% of requests on August 5 and reaching 16% in the worst hour on August 31. Anthropic said about 30% of Claude Code users who made requests in that stretch had at least one message routed to the wrong servers. People who said Claude had gotten worse that summer were right, and the cause was infrastructure bugs, not a quiet downgrade.


 

Five things that can make Opus 5.5 feel worse

The effort level is lower than you think. Anthropic's model docs list Opus 5.5's default effort as Medium, one level below Opus 5's High. Claude Code also starts Opus 5.5 at Medium. If you moved from Opus 5 without touching the setting, the same prompt now gets less thinking. Run /effort status to see the current level. The Claude Code docs also note that effort is calibrated per model, so High on Opus 5.5 isn't the same amount of work as High on an older model.

Your session switched models without you noticing. Opus 5.5 runs safety classifiers that mostly flag cybersecurity and biology content. When a request trips one, Claude Code re-runs it on an older model, Opus 4.8 for cyber or Opus 5 for biology, and posts a notice in the transcript. The session then stays on that older model until you change it back with /model. One flagged question early in a session can leave you on Opus 4.8 for the rest of the afternoon.

"Opus" isn't the same model everywhere. In Claude Code, the opus alias means Opus 5.5 on Anthropic's own API, but the docs say it resolves to Opus 4.6 on Microsoft Foundry. The opusplan setting uses Opus only in plan mode and Sonnet for the actual work, and an ANTHROPIC_MODEL variable left in your shell overrides whatever your settings file says. Run /status to see the model you're really on.

The conversation is too long. As a session approaches its context limit, Claude Code compacts it, summarizing the earlier conversation to make room. Details from the start of a long session can get lost in that summary, and Claude can lose track of things you said early on. The /context command, one of Claude Code's built-in commands, shows how full the context is. Starting a fresh session for a new task often fixes the "it got dumber" feeling outright.

Something is actually broken. Check status.claude.com for incidents. Anthropic's 2025 postmortem showed that problems can hit some requests and not others, which is why user reports can contradict each other for days.

Your own feelings are a bad benchmark, because the same prompt gives different answers on different runs anyway. Keep two or three real tasks you know the right answer to, like a bug you fixed last month or a function with tests. When you suspect a downgrade, rerun them in a fresh session with the model pinned by its full name, claude-opus-5-5, and a fixed effort level, and run each one a few times. If the results match what you got before, the model is fine and something in your usual setup changed. If they're clearly worse, use /bug to send Anthropic the conversation, since a report with a real transcript is far more useful than a forum post.

Some changes are on purpose. Newer Claude models act more on their own and write longer answers, which a lot of people read as worse. We covered that in why Claude can feel worse while scoring better, and our first take on Opus 5.5 explains why many people liked it more than Opus 5. For picking the right effort level, see the effort levels guide.


 

Sources

Anthropic: A postmortem of three recent issues - The statement that Anthropic never reduces quality for demand or load, the three 2025 infrastructure bugs, their dates, and the share of requests and Claude Code users affected.

Claude API docs: Models overview - Opus 5.5's default effort of Medium.

Claude Code docs: Model configuration - Default effort in Claude Code, per-model calibration, aliases by provider, opusplan, ANTHROPIC_MODEL, and automatic model fallback after safety classifier flags.

Claude Code docs: Commands - /effort, /status, /model, /context, /compact, and /bug.

Claude status page - Current and past incidents.

Found this article useful?

Add ClaudeFolio as a preferred source on Google to see our articles first.

FAQ

Has Anthropic nerfed Opus 5.5?
There's no evidence of a deliberate downgrade, and Anthropic says it never reduces quality because of demand or server load. Most complaints trace to lower default effort, a silent fallback to an older model, or very long sessions.
Why does Claude Code feel dumber in a long session?
As a session nears its context limit, Claude Code compacts the conversation into a summary, and details from early on can get lost. Check usage with /context, and start a fresh session for a new task.
How do I check which model Claude Code is using?
Run /status to see the current model. If a safety classifier flagged a request, Claude Code may have switched the session to Opus 4.8 or Opus 5, and you can switch back with /model.

Related posts

Comments