If your Claude Opus 5 or Opus 5.5 chat suddenly answers with an older model, a safety check flagged the request and Claude fell back to a less capable model, or blocked it. As of October 2026 automatic switching is on by default, Anthropic says most requests never hit it, and you can switch back from the model picker or turn the switching off in Settings > Capabilities.
The catch: the checks read more than your message
This is what catches people out. Opus 5 and Opus 5.5 run automated safety checks, which Anthropic calls classifiers, on every user request. Those checks don't stop at your latest message. They review everything the model reads, and that includes memory, content pulled in from connectors, web search results and files.
So a fallback can be triggered by something you didn't type. A document you attached for a quick summary, a search result Claude fetched, or something sitting in memory from an older chat can all be the reason the model switched. If the message you just sent looks harmless, look at what else is in the conversation.
Anthropic's framing is that a narrow set of higher-risk requests fall back so that everyday work can carry on while misuse is limited. If you want the wider picture of what Opus 5.5 is and how to use it day to day, my guide to using Claude Opus 5.5 covers that; this page is only about the fallback.
The four classifiers and what they do
There are four classifier types, and they don't all behave the same way. Here's what trips each one and what happens, as of October 2026.
| Classifier | What trips it | What happens |
|---|---|---|
| Cybersecurity | Offensive requests: exploit generation, binary vulnerability scanning, penetration testing | Opus 5 or 5.5 to Opus 4.8 |
| Biology | Dual-use virology, toxicology, molecular design (Opus 5.5 only) | Opus 5.5 to Opus 5 |
| Frontier LLM development | Certain accelerator kernel work (Opus 5.5 only) | Opus 5.5 to Opus 5 |
| Distillation | Asking for verbatim reasoning or full chain of thought | Blocked outright, no fallback |
Cybersecurity. Defensive work still runs on Opus. Secure coding, scanning source code for vulnerabilities, triaging issues and building secure code all work. The fallback is for offensive requests such as writing exploits, binary-based vulnerability scanning and penetration testing.
Biology. Opus 5 doesn't fall back on biology, chemistry or life sciences, and uses safeguards similar to Opus 4.8. Opus 5.5 is stricter, with classifiers similar to Fable 5, and falls back to Opus 5 on dual-use requests in virology, toxicology and molecular design. Everyday health and education questions still work.
Frontier LLM development. This only applies to Opus 5.5. A small set of capabilities, such as kernel development for certain ML accelerators, sends it back to Opus 5. Anthropic says it shouldn't affect the vast majority of ordinary AI or ML development, research or general coding.
Distillation. These checks block attempts to pull out the model's internal reasoning, like asking Claude to repeat its reasoning word for word or write its full chain of thought to an external output. These don't fall back; they're blocked. You can still ask Claude to explain its reasoning, teach a concept or walk through a code review, and asking 'why did you do that?' is fine.
What happens after a fallback
With automatic switching on, Claude re-runs the blocked request on a less capable model in the same conversation. You'll see a notice, and the answer is labelled with the model that wrote it. The model picker then stays on that less capable model for the rest of the conversation, which is why later replies can feel weaker even when your new questions are harmless.
You can switch back to Opus 5 or 5.5 whenever you like. The same safeguards may trigger again if the original request is still in the conversation, though, and Anthropic says editing your previous message before retrying often helps.
How to get back to Opus
- Look for the notice saying the model switched and the label on the answer.
- Edit the message that triggered it, or remove risky content from memory, files or connectors.
- Switch back to Opus 5 or 5.5 from the model picker.
- If you prefer a pause over a switch, toggle off 'Switch models when a message is flagged' in Settings > Capabilities.
- For legitimate security or life-sciences work, apply to the relevant verification program.
In Claude Code the same toggle lives under Config > MODEL & OUTPUT. With it off, a request that would fall back pauses the conversation instead. You can then edit and retry on Opus, or send the same message to a less capable model yourself.
If the fallback model blocks it too, that's because Opus 4.8 has its own safety systems; edit the message and try again. For legitimate defensive security work there's the Cyber Verification Program, and life sciences research organisations can apply to the Life Sciences Verification Program. Opus 5.5 isn't currently available in the Cyber Verification Program, but organisations already using Opus 4.8 through it can use Opus 5 with fewer cyber restrictions, and Opus 5 is compatible with Zero Data Retention.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Are blocked requests billed?
Sometimes, and it depends on the classifier. As of October 2026, refusals that arrive before any output are billed when they stop or fall back because of the biology, distillation or frontier LLM development classifiers. Blocks before output in other categories aren't charged.
If a request is blocked midstream, the input tokens and the tokens already streamed are charged at the rate of the model that produced them. A fallback response is charged separately at the responding model's rates, with a credit for the cache miss.
Where fallbacks apply
Fallbacks apply across Claude on the web, Mobile, Desktop, Cowork, Claude Code, Claude Design, Claude for Microsoft 365, Claude Tag and Claude Science. The Claude API is the exception. Switching isn't active there by default: developers have to opt in and configure fallbacks, and until they do, the model returns a 200 response with a stop reason. If you build on the API, handle that stop reason or your app will look like it simply went quiet.
Effort and thinking on Opus 5 and 5.5
The model menu next to the send button sets three things: the model, the effort level and thinking. Effort runs from Low and Medium for routine tasks, through High for the best balance, to Extra high for long-running coding and agentic tasks and Max for the deepest reasoning. Higher effort uses more tokens, so you'll reach your limits faster.
Thinking can't be turned off in Claude for Opus 5 or Opus 5.5, and on Opus 5.5 it's always on at every effort level, including on the API. On the API, Opus 5 can have thinking disabled at effort high and below, but disabling it at xhigh or max returns an error. Changes apply from Claude's next response, and on Enterprise plans an administrator may have turned a model or effort level off for your role. For more on how thinking works, see my extended thinking explainer.
My verdict
I'd leave automatic switching on if you mostly write, research and code, because Anthropic says most requests never trigger it and a quiet switch beats a stalled chat. I'd turn it off if you work near the edges, such as security or life sciences, and you'd rather edit and retry on Opus than carry on with a weaker model without noticing. Either way, when a switch happens, check your files, memory and connectors before you blame your own prompt.
Frequently asked questions
Why did Claude switch from Opus 5 to another model?
A safety classifier flagged the request, so Claude re-ran it on a less capable model. The checks read memory, connectors, search results and files as well as your message.
How do I stop Claude switching models automatically?
Turn off 'Switch models when a message is flagged' under Settings, then Capabilities. In Claude Code it's under Config, then MODEL and OUTPUT.
Can I switch back to Opus after a fallback?
Yes, from the model picker at any time. Editing the message that triggered it before retrying often helps.
Does the Claude API fall back automatically?
No. As of October 2026 developers must opt in and configure fallbacks; until then the model returns a 200 response with a stop reason.
