Claude Code "Opus 5.5's safeguards stopped the response above · continuing once with that noted"

Why it happens

Anthropic runs safeguards classifiers on some models. When one flags a request, the response is stopped and Claude Code applies its automatic model fallback (Claude Code 2.1.219 or later): it re-runs the request once on a fallback model and keeps the rest of the session on that model.

1. The request touched a flagged area. [cyber] is the one most developers hit: security tooling, exploit code, scanning or malware analysis, even when the work is defensive.
2. False positives. The in-product message says Anthropic is improving these safeguards to reduce incorrectly flagged messages. Two open reports from 9 November 2026 (anthropics/claude-code#108497 and #108506) show ordinary sessions being stopped turn after turn.
3. The session stayed on the fallback model. After a fallback, later turns run on Opus 4.8 (for cyber) or Opus 5 (for biology) until you switch back, which can feel like the model got worse.
4. A fallback bug fixed in 2.1.321. Its changelog says it fixed the fallback model sometimes declining a request or mentioning a stopped response when the first model was blocked before replying.

The fix

1. 1 Read the Details line to see which category fired.
2. 2 Rephrase the task or split the flagged step into its own request with clear context, and keep the rest of the work in the main session.
3. 3 Return to your model with /model, or start a new session, which starts on your default model again.
4. 4 Check which model actually answered: /cost reports usage per model. For a turn-by-turn record, assistant entries in ~/.claude/projects/<project>/<session-id>.jsonl carry the model name, though Anthropic notes that format is internal and changes between versions.
5. 5 Update Claude Code (claude update). Versions before 2.1.321 had the fallback bug above.
6. 6 If you would rather choose than be switched, the model configuration docs describe an "Ask each time" option (switchModelsOnFlag) and a way to turn the fallback off with CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1. With it off, a flagged turn simply ends with the error.

Which model does Claude Code switch to?

Per Claude Code's model configuration docs: on Fable models and Opus 5.5, cybersecurity-flagged requests re-run on Opus 4.8 and biology-flagged requests on Opus 5. Sonnet 5.5 cyber flags fall back to Sonnet 5.

The switch is not just for the one reply. The docs say the session continues on the fallback model afterwards, so check your model before judging the quality of later answers.

Still failing?

• Every turn ends with "API Error: ... safeguards flagged this session". The fallback is off or failed. Start a new session and rephrase, or pick another model with /model.
• Your sessions feel slower or weaker without any safeguards line. Check auto compact, context size and effort level before blaming the model.
• "Prompt is too long" instead. That is a context limit, see the page below.

Related errors

• Claude Code "Prompt is too long"
• Claude API error 529 overloaded
• Claude Code "expected thinking block"