Claude switched to an earlier model: why it happens and what to do

If you use Claude Opus 5.5 and at some point the notice "passed to an earlier model" appears Switched to [older model]), is not a failure and is not a limit of your plan. It is the expected behavior of Safety Filters: Opus 5.5 is the first Opus model released with bio and cyber protections at the same level as Fable, and when a message is reported the request switches to an earlier model.

The practical problem is that change can happen on entirely legitimate work — Anthropic says it openly and is taring the filters — and that from that moment answers a different model. This guide explains why this happens, how to go back in the app and in Claude Code, and how to reduce reports. 23 September 2026 on the official guide.

Why It Happens

Opus 5.5 comes with the strictest filters ever applied to an Opus model: bio and cyber protections at Fable level. The check does not look only at the text you write: it also covers attached files and the results of searches the model performs for you. When something is flagged, the request is not blocked: it is served by an earlier model, and the interface says so.

Anthropic states that this produces false positives on legitimate work and that calibration is ongoing. Some categories are explicitly allowed: search for vulnerabilities in the source code, health and educational questions.

What is NOT this warning

It is worth saying, because the first reactions are almost always wrong.

It is not a block of your account. No plan limitation, no suspension: the next message returns to normal.

It's not a judgment on the content. The filter works by categories and patterns, it doesn't read intentions. A security audit on your software can be reported exactly as a problematic request, and for this there is false positives reporting.

It's not a technical error. No need to reload, change browsers or restart Claude Code. Behavior is expected and documented in the official model guide.

It's not permanent. It applies to that request. In the app, if the automatic transmission is active, subsequent requests are re-evaluated one by one.

What to do in the Claude app

1. Read the notice. It says which model it's responding to. If you're doing a long job, it's time to notice.
2. Back to Opus 5.5 from the model selector, high in conversation.
3. If you prefer that you never change on your own: Settings → Features → disable "Switch models when a message is flagged". From that point on, a flagged message is not served by another model.
4. Reformula the message that caused the report (see below the most common causes).

What to do in Claude Code

1. The notice appears with the name of previous model who took charge of the request.
2. /model back to Opus 5.5.
3. Esc twice Allows you to edit the message that caused the report.
4. /config contains the option "Change model when a message is flagged", to be deactivated if you do not want the automatic switch.
5. /feedback signals a false positive to Anthropic. It really serves: that's how the calibration improves.

For the rest of the workflow with Opus 5.5 in Claude Code — long runs, stop rules, subagent — see Opus 5.5 in Claude Code.

The Most Common Causes, and How to Avoid Them

The first line is the most frequent and the easiest to delete: ask the model to reproduce the internal reasoning is a reported category, and in most cases it serves no purpose. A motivation in three sentences gives the same transparency. The theme of the prompts to be reviewed is in prompt for Claude Opus 5.5.

How to tell if it's a false positive

Before rewording, three questions:

• The message asked to show the internal reasoning? If so, that's almost certainly it: rephrase the request for motivation and try again.
• Was there a long attachment? Try the same message without an attachment. If it passes, the problem is in the file: attach only the necessary part.
• Did the model do a web search? Repeat the question by asking not to search. If it passes, one of the pages found has the filter activated.

If all three answers are no and the request falls into an admitted use —your code, health, training —it is a false positive to report with /feedback in Claude Code. The more reports arrive, the earlier the calibration gets better: Anthropic explicitly declares it as work in progress.

When to leave the automatic transmission active

It is not said that disabling it is the right choice for everyone.

Team and Enterprise: who decides the setting

In group plans the question "Do I leave the automatic transmission active?" does not have a single answer, because it depends on the work of each. Three practical indications.

Do not impose a single choice on everyone. Those who do long analyses and those who use chat to respond to emails have opposite needs: the first always wants to know which model answers, the second always wants an answer. Let each person choose, but explain what the notice means.

Keep a record of recurring false positives. If the same kind of legitimate request is reported to multiple people capitait happens on teams working on cyber security or health issues —it is worth collecting the examples and sending them all with /feedback: a documented report weighs more than ten complaints.

For automated systems via API, a different discourse applies. There is no app that changes its model on its own: the response to a reported request follows the rules of the platform, and must be managed in the code like any other unexpected outcome. If you have automations in production, check the official documentation how reports are returned and predict an alternative route, as explained in AI Agents for Companies.

Practical examples

Software house, 15 people, Turin. During a security audit on your code the notice appeared on each file. The message was correct —finding vulnerabilities in its source is an admitted use —but each time the response came from an earlier model, less precise in false alarms. Solution: Disable the automatic transmission in /config, reformulate the aperture ("we're checking our code before release") and report false positives with /feedback.

Associate Medical Study, Bologna. Clinical questions to prepare information material for patients. Health questions are allowed, but a long annex with pharmacological terminology triggered the alert. Solution: Attach only the necessary sections and state the purpose in the message. No settings touched.

Mistakes to Avoid

• Ignoring the warning. The model change is silent for the rest: if you don't read it, you work with a different model without knowing it.
• Insisting with the same message. Rewording works, repeat no.
• Turning off the automatic transmission and then marvel at not having an answer. With the option off, the flagged message is not served by others.
• Not reporting false positives. /feedback is the channel with which calibration improves.
• Asking for internal reasoning "for safety". It's the quickest way to get flagged for no reason.

How to apply it in the company

If multiple people in the company use Claude, three rules to put in writing:

1. Who works on long tasks Turn off the automatic transmission and check the model before you start.
2. Who Uses Chat for Various Work He leaves it active and learns to read the warning.
3. No one attaches whole documents when one part is sufficient: reduces reports and protects data, as per rules on business data and AI.

For the complete picture of the model — prices, benchmarks, availability — Claude Opus 5.5: what is it.

Conclusion

The "passed to an earlier model" notice is a consequence of filters, not a mistake: Opus 5.5 protects more and sometimes mistakes for excess. Knowing that it exists is already half the work. The other half is one rule: read the warning, reformulate, report if it is a false positive. The settings are there, in the app and in Claude Code, to choose the behavior you prefer.

AI becomes useful when it enters processes. Giallo Studio helps SMEs and teams to build concrete, measurable and sustainable automations.