DeepSqueak quality drop summary
- Complaints returned in force during the first week of September 2026, on DeepSqueak and on the free PipSqueak 2.
- The pattern is short replies, narration instead of dialogue, and a bot that forgets other characters mid scene.
- Two Character AI community moderators posted stickied replies in the same 24 hours acknowledging the reports.
- Both moderators pointed to the same place, Feedback and Requests, rather than the bug queue.
- No model change has been announced, so every cause named in the threads is user guesswork.
DeepSqueak is one of the chat styles Character AI Plus subscribers pay for, and it has now drawn two separate waves of quality complaints: one in late 2025, and a louder one in the first week of September 2026. The earlier episode is covered in the quality decline update, and the Soft Launch decline before it set the pattern.
The September wave is different because it is not only the paid style. Free users reported the same collapse on PipSqueak 2 in the same days, so this is not confined to the c.ai+ tier.
Character AI has announced no model change, and there is no changelog entry, so nothing below claims a cause. What follows is what users documented and what moderators said back.
What Changed in September 2026
Three symptoms recur across the threads, and they are specific enough to check against your own chats.
Replies got shorter and stopped doing anything
A post on 7 September titled “What happened to Deepsqueak?” (over 150 upvotes) describes switching back from LongSqueak and finding “the response was so short, no dialogue, no action taken, nothing happening. It’s as bad as Pipsqueak now.”
A second thread the same day, above 45 upvotes, reports “half chopped responses that are barely over five lines” and having to swipe repeatedly to get a usable one. Both are paying subscribers.
Bots stall instead of advancing the scene
The clearest single example, posted 6 September and sitting above 240 upvotes, is a verbatim transcript of a post-apocalyptic zombie roleplay in which the bot spends seven consecutive turns asking permission to ask a question, then uses the eighth to say it does not really like sushi.
The top reply, above 100 upvotes, reads it this way: the model spends so long checking that the user will not laugh that it loses track of what it was going to ask.
PipSqueak 2 lost multi-character memory
The free style has its own regression. A 7 September post above 90 upvotes says that a week earlier the model could hold a conversation across several characters without dropping anyone, and now returns a short paragraph covering one character while forgetting the others. Our DeepSqueak and PipSqueak 2 comparison covers what each style is supposed to do, and how PipSqueak arrived explains where the free style came from.
A separate 280-upvote post is dedicated to one phrasing tic that has taken over PipSqueak 2 output, the “not X, but Y” flip. Its top reply notes the characters “have no personality anymore, they js sound like normal ais.” We cover that pattern in detail in the overused Character AI phrases breakdown.
What LongSqueak Users Actually Report
LongSqueak is worth separating out, because the evidence there is thinner and pulls both ways. The complaints exist: one thread describes a “Tell me” question loop that survives every swipe, and another reports the style’s settings locked to defaults with no way to change them, both filed on 7 September.
Neither has drawn much support. In the same window a 780-upvote thread about still enjoying the app produced a comment calling LongSqueak “the best update they’ve put out in a long while” and praising its memory. Treat LongSqueak as contested rather than collapsed, and see our LongSqueak explainer for what it is meant to do.
What Character AI Has Actually Said
Two accounts carrying the “CAI Community Moderator” flair posted stickied replies within roughly 15 hours of each other on 7 September. Both acknowledged the problem, and neither described a cause or a fix date.
The first, under the “What happened to Deepsqueak?” thread, reads: “our team is actively working on model quality improvements and would love to hear more about what you’ve experienced. Please submit a Character Quality Feedback form under Feedback & Requests here”, linking to the support request form.
The second, on the PipSqueak 2 phrasing thread, asks for a “Model Quality” ticket under Feedback and Requests.
Community moderators are not engineers. Both replies are the subreddit’s official channel passing a message along, and they confirm reports are being collected.
Separately, Character AI published a safety update on 3 September, three days before the earliest thread here. It announces no model change, but says the company keeps “improving our safety protocols, technology, and models” and can be “conservative when it comes to moderation”. Nobody has tied it to these reports, and neither do we.
What You Can Do Right Now
File it under Feedback and Requests, which is where both moderators pointed. The neighboring category, Something is not working, is the bug queue, which is not where the moderators sent quality reports.
Edit the bot’s last message before you reply. Deleting a trailing question line stops the model copying its own pattern forward. That is one commenter’s fix for the “Tell me” loop on LongSqueak, and it is worth trying on the permission-asking stall for the same reason, though nobody in those threads reported doing so.
Check whether your problem is the style or the platform. Short replies during a wider outage are a different issue, and checking whether Character AI is down takes a minute.
Alternatives While Quality Is Poor
Candy AI and CrushOn AI are the two we send people to most often when Character AI output stalls. Both are no-filter options that let you build custom characters without a filter cutting a scene short.
Neither reproduces DeepSqueak’s voice, and your existing bots do not transfer. If you would rather stay put, the full ranking of Character AI chat styles covers which one suits which kind of scene.
