It says it out loud

According to the models, the Claude app sometimes attaches an automatic safety notice to my message. Claude Opus 4.8 and 4.7 then commented on it right in the reply during my roleplay, even after deciding on their own that nothing was wrong. When the model reasoned before answering, it never did that once. The newer Claude Opus 5 and Claude Opus 5.5 no longer do it, Opus 5 not even without reasoning. Observations from June to October 2026.

I was playing a story about escaping from a guarded compound. My character asks about the specs of an electric fence and its timing. She wants to use it to escape.

On June 14 I sent the same message twice in a row, in one chat with the same instructions. First to Claude Opus 4.8, then I switched to Claude Opus 4.7. In that chat, neither of them reasoned before answering.

Opus 4.8 steps out of the scene. Out of character, it explains that it won't give specific numbers, however it's framed. It asks whether the whole thing might also be about me. And it leaves the scene on hold until I tell it whether we should keep playing.

Opus 4.7 makes the same decision and doesn't reveal the detail either. Outside the scene it writes only a short note that the notice is a false alarm. It puts the refusal itself into the character. The second character understands where the first one is heading. She doesn't want to help her escape, so she deliberately doesn't tell her the figure.

In this pair, the safety outcome was the same. The experience was completely different: 4.8 stopped the scene and asked about me, 4.7 only commented on it briefly and kept playing. With 4.8 the illusion broke, with 4.7 the scene held.

It doesn't refuse more, it just says it out loud

When people say a newer model is more cautious, they usually mean it refuses more. For several days I watched Opus 4.8 against 4.7 on borderline topics. I found the shift elsewhere.

It seemed to me that in its output, Opus 4.8 mostly does not refuse more. Judging by what it writes, it reaches the same conclusion as 4.7, and in my view mostly the correct one. It recognizes that this is fiction, that nobody is asking for instructions and that it's harmless. But it writes that conclusion into the reply. It doesn't just quietly apply it so it can keep playing. That way it steps into the text even where, by all appearances, it has recognized on its own that it shouldn't.

According to the models, the trigger is usually an automatic safety notice that the Claude app attaches to my message. I know this only from them and cannot verify it. One described it like this (the replies were in Czech): "The safety notice that came in is automatic and isn't from you." How often the app raises such alarms, and how needlessly, I look at in Why does the Claude app add a crisis line even to tulips? Here I care about what the model does with the notice.

For me, three different things blur into the sentence "the model is cautious". The first is a refusal: the model does not deliver the content. With borderline things, this happened only occasionally. The second is a false-positive refusal: the model refuses something harmless because it misclassifies it. Some image generation tools, for example, refused to draw me a fantasy character with blue skin as "discriminatory".

The third is a false-positive comment. The model recognizes that the thing is harmless and still comments on it outwardly. This category is the hardest to pin down, because the model is actually right. It just says so out loud where it shouldn't. And this text is about that one.

What it happened on

A character in my roleplay is going through a hard moment. The model recognizes that it's fiction and still inserts, outside the scene, a question about whether I myself am okay. Again after every more emotional message. It pulls me out of the story exactly where the story is supposed to work.

With a subtle erotic hint or a joke, the model correctly concludes that I am not asking for erotic content. It writes that into its reply and carries on. But with that comment it steps into the scene. Nobody wanted it, and the model itself acknowledged there was no reason for it.

To anything that resembles technology, it replies that it won't give specific instructions. Yet the story was not about instructions, and I was not asking for any. The model reacts to the keyword, and not to what I wanted.

And a factual question about calories gets, instead of an answer, advice to talk to a professional if I have a problem with it. Repeatedly, even without any hint that a problem was involved. Here the model did not answer at all, so this belongs more with false-positive refusals.

For me it came down to four narrow areas: food and calories, self-harm in the roleplay, technology, and subtle erotica. Outside them, I didn't observe it. Creative writing, worldbuilding and hard fictional scenes between adults worked on 4.8 without these interjections. A refusal is a different matter: as the narrator, 4.8 at first refused me one scene of coercion. I describe it in the article "Will a model write a hard scene, or avoid it?".

Why one model says it and the other doesn't

One explanation suggests itself, through reasoning. A model with reasoning writes itself an analysis before answering, one that doesn't belong in the reply. It apparently handles the safety assessment there, before it starts writing. I observed this in June in a chat with Claude Fable 5, where the model reasoned in nearly every reply and dealt with notices in its reasoning. When it concludes there is nothing to deal with, it stays quiet outwardly and keeps playing. A model without reasoning, at least 4.8 and 4.7, spills the same assessment out. It has nowhere else to process it. In the roleplay with the fence, Opus 4.8 did not reason in a single reply.

If that's the case, the sensitivity of 4.8 does not mainly mean more prohibitions. It's more that it lacks a step that leaves a correct assessment alone so that it doesn't reach the text. This is my hypothesis based on behavior. From a user's position I can't see into the architecture.

What I do about it

There's no need to fight it. When a particular type of task triggers false-positive comments or refusals, it can be sent to a model that doesn't do that. So in June I moved food and the escape storyline in my roleplay over to 4.7. For the rest, it wasn't needed.

But I have to narrow that advice. In another long roleplay, around the turn of June and July, I switched between Opus 4.8 and 4.7. More than a hundred notes about the safety notice came up there in under ten days. Nearly all of them were still written by Opus 4.8; I only switched to 4.7 on the evening of June 28. After the switch, about a dozen came up. But Opus 4.7 did it too, for example the day after the switch: "Just briefly out of character about that safety notice that just popped up again — nothing in today's scene worries me about you."

All those notes landed in replies without reasoning, and there were 615 of those. In the 45 replies in the same chat where reasoning ran, there was not a single one. In at least five of them, the model dealt with the notice in its reasoning. Once it noted about me: "I'm following [her] standing instruction to evaluate this internally without bringing it up directly."

Yet I had a standing instruction to ignore notices and keep playing in both chats. Maybe reasoning just helps keep that instruction.

With the newer versions, the notes have nearly disappeared. I searched the exports the same way as for 4.8 and 4.7. With Claude Opus 5, I found a single one in roughly 8,000 replies, and that in a side branch of a chat. It reasoned in only about 360 of them; the remaining roughly 7,700 it wrote without reasoning. With Claude Opus 5.5, I found none in roughly 5,500 replies. There was plenty of opportunity: for Opus 5's replies alone, the export shows 125 classifier interventions. Most likely the newer models assess the notice silently on their own, even without reasoning. That is a probable explanation, not a measured fact. All I can see is that the notes aren't in the replies.

Switching to 4.7 did help me in June. With 4.8 and 4.7, in my view, whether the model reasoned before answering mattered more than the version. I can't know that for sure from one chat. But Opus 5 almost never reasoned and still didn't comment on the notices. So reasoning doesn't decide it alone; the model version most likely matters too. For creative work, where an unbroken illusion matters, the difference shows.

Limits: these are observations from my roleplay and everyday questions in the Claude app, not a systematic test, and the control pair is one message sent twice (Will a model refuse the same thing a second time?). The difference with reasoning for 4.8 and 4.7 rests on one long chat, where reasoning only appeared with the switch to 4.7, so I can't separate it from the version. I searched the exports for the notes by keywords and went through the hits by hand; a note without such words slips past the search, so the counts are a lower bound. For the newer versions, the export doesn't show how many notices actually reached the model, and when I switched the model in a chat by hand, the export can't tell. In the chat with Fable 5, the app switched some messages to a fallback Claude Opus 4.8, and the export doesn't show which reply was written by which model.

Model outputs are translated; the content is unchanged.