Claude Code Switched From Opus 5.5 to Opus 4.8? Why, and What to Do

スポンサーリンク
Claude Code Switched From Opus 5.5 to Opus 4.8? Why, and What to Do Claude
スポンサーリンク

I was discussing some work with Claude Code when its reply stopped halfway, followed by a notice saying it had switched the model to Opus 4.8. My Claude Code runs in Japanese; translated, the notice said:

Switched to Opus 4.8
Opus 5.5's safeguards flagged this session. This may be the first time you're seeing this notice with an Opus model. Opus 5.5 is more capable, with correspondingly stronger safeguards, so it can sometimes flag work that isn't cybersecurity. We're improving the safeguards to reduce messages flagged by mistake.
Details: [cyber]
The notice in Claude Code (Japanese UI) saying Opus 5.5 safeguards flagged this session
The notice as shown in the Japanese UI

I wasn’t asking about hacking anything. I checked how this works in Anthropic’s support article, so here is the mechanism along with what actually happened.

Short answer: Opus 5.5 can hand a session to Opus 4.8 when a safety classifier flags it as cybersecurity. You can switch back with /model, but if you keep the same topic it may switch again. Changing the topic, or starting a new conversation, worked for me.

What I was doing when it appeared

I was asking Claude Code how to reduce the permission prompts in the Claude Code I use at work. My company’s setup disables auto mode (letting the AI handle approvals), so when I run several sessions in parallel, prompts pop up everywhere. I asked whether there was any way to make that more efficient without auto mode.

Partway through, Claude’s replies started getting cut off as “stopped by a safety classifier”, then the notice above appeared, with the model switched to Opus 4.8. The details said [cyber] (cybersecurity). I think the topic of “doing something about the permission system” was judged to be a security question.

What happened next:

  1. I switched back to Opus 5.5 with /model, then continued the same topic (automating approvals). After a short exchange it switched to Opus 4.8 again
  2. I switched back to Opus 5.5 once more, gave up on the automation idea, then moved on to writing this blog post. Since then it has stayed on Opus 5.5

What’s going on (the official explanation)

Anthropic’s support article “Why Claude switched models in your conversation with Opus 5 or Opus 5.5” describes the mechanism. The key points:

  • Opus 5 plus Opus 5.5 check every request with automatic safety classifiers. Not just your messages, but also memory Claude reads, connector data, search results, file contents
  • The areas covered: cybersecurity, biology, frontier LLM development (Opus 5.5 only), distillation
  • When the cybersecurity classifier flags something on Opus 5.5, it may switch to Opus 4.8
  • After switching, it doesn’t go back to Opus 5.5 by itself; you have to switch back
  • You can switch back to Opus 5.5 from the model picker at any time. But if the flagged message is still in the conversation, it may switch again

That matches what I saw. The first time I switched back while staying on the same topic, it switched again. The second time, after changing the topic, it stayed on Opus 5.5.

What to do when it happens

1. Just keep going on Opus 4.8

Opus 4.8 is plenty for everyday work. If the topic is still security-adjacent, not switching back is the least hassle.

2. Switch back to Opus 5.5

In Claude Code, use /model:

/model claude-opus-5-5

If you continue the same topic afterwards, it may switch again. For me it recurred on the same topic, but not after I changed the topic.

3. Rephrase the message, or start a new conversation

The support article says rephrasing the earlier message before retrying often works. If you’re moving on to unrelated work, a new conversation (session) avoids dragging the flagged message along.

4. Stop it from switching automatically

In Claude Code, turning off “Switch models when a message is flagged” under “MODEL & OUTPUT” in Config makes it pause the conversation instead of switching models when flagged (per the support article). Use this if you don’t want the model changing behind your back.

5. Send feedback if it’s a false positive

If unrelated work gets flagged, you can report it as a mistake via “Send feedback”. The notice itself says they’re working to reduce mistaken flags, so reports should help.

My take

My impression is that it wasn’t about going against my company’s managed settings. Asking Claude Code for a way around its own mechanisms was probably mistaken for a security question.

When I once got stopped by reasoning_extraction, that also appeared without me intending anything. When you hit one of Anthropic’s guardrails like this, it seems best not to push: change the topic, or start a new thread (conversation).

For the record, the only thing I do to reduce permission prompts in my work setup is use acceptEdits mode, which lets file edits through automatically.

Summary

  • Opus 5.5 safety-checks every request; when something is judged as cybersecurity it may switch to Opus 4.8
  • It stays on Opus 4.8 until you switch back. /model brings it back, but the same topic can trigger it again
  • For different work, use a new conversation. To keep the model from changing, turn off “Switch models when a message is flagged”
  • If it’s a false positive, report it with Send feedback

Related posts

References

コメント

Copied title and URL