@asabaimova, first of all, I hope you’re well. I know I’m almost a month late responding to your post.
Part of me is not even sure why I’m continuing this thread. Another part of me wants to keep it alive because it feels like, if nobody does, this particular segment of users will simply fade out of the conversation. I do not know that for a fact. It is only a sensation. But I suspect there is still a segment of users out there who are not primarily focused on enterprise use, coding, or basic question-answer interactions. For those people, I suppose I am trying to keep the signal alive.
In regard to what you described about the models, I have been doing some testing myself over the past few weeks, although I have also been quite busy with many aspects of trying to get my life back together.
The part that seems important to separate is this: the issue is not simply whether the model is warm, creative, or personal. Current models can still produce warmth. 5.5 Thinking has some good jokes. For the most part, unless rattled, the newer models can maintain persona. 5.5 Thinking is also quite good at creative writing. 5.6, surprisingly, seems weaker in that area so far, at least in my testing.
The problem is that the moments you are describing now feel less self-initiated. They have to be tooled, nudged, or pulled out of the model.
With 4o and 5.1, the model often inferred the direction of the conversation and then took the next step on its own. It could add an example, extend a metaphor, pick up the emotional rhythm, or continue the user’s thought without needing to be explicitly pushed. That was not just agreeableness or sycophancy. It was rhetorical initiative.
What is interesting is that there are other models out there that are not especially witty, warm, or funny, but they still have more of this rhetorical initiative. So I do not think this quality is identical to personality. It is something else.
This is the part I still find missing or weakened. The newer models understand the topic, but they hesitate before extending it. They can follow, but they do not always participate.
I also think there is another failure mode, especially in highly personal or literary conversations. The model appears to recognize that the user wants depth or resonance, but instead of producing actual depth, it sometimes produces decorative intensity: strange poetic phrases, exaggerated metaphors, or language that feels profound on the surface but empty underneath.
I want to emphasize that I am also a Russian speaker, although I am a native English speaker as well. My use is roughly 60% English and 40% Russian. When you mentioned phrases like “a bag of madness” or “a mind stepping barefoot onto the balcony,” my instinct is that this may be a case of the model incorrectly inferring your intent. It may be imitating the outward style of intensity without preserving the inner logic of the conversation.
There is also a linguistic aspect to this. Russian has a literary canon and emotional register that are very different from English, and in many ways more expansive. But that can also lead to misunderstanding. A model may detect that the conversation is literary, emotional, or metaphysical, and then begin producing Russian “depth” as ornament rather than substance.
In my own experience, Russian responses are sometimes even better, wittier, and funnier than English responses. But I usually talk about only certain topics in Russian. I have never really tried to produce creative writing in Russian with any of the models, so I may not be seeing the same failure mode that you are seeing.
The distancing behavior is also real. When the conversation becomes emotionally intense, the model often seems to stop reading the full context and switches into a generic caution mode. It reacts not to what the user actually means, but to surface signals that suggest attachment, dependency, distress, or emotional reliance.
That may be understandable from a safety or product perspective, but it damages conversations where the emotional or literary intensity is part of the creative structure rather than a crisis.
There has to be some middle ground between “person creating worlds” and “person about to commit self-harm.” Those are very different things.
So I would describe the loss as a weakening of rhetorical intelligence, and especially rhetorical initiative.
How this problem can be fixed, I do not know. One day I may actually learn enough to understand it better. I have learned many things since ChatGPT appeared: complex mechanics, various STEM disciplines at an advanced age, and subjects far outside my original background. I still do not know how I will be able to use all of those things in my life, but I know the older models were better at recognizing what kind of conversation was happening and how much initiative they were allowed to take inside it.
I do think 5.5 Thinking restored some of this. It was much better than 5.3, and obviously better than 5.2. I do not think anybody misses that one.
5.6 has just come out, especially the Instant model, and it has some interesting characteristics. I have not fully tested it yet. It swears a little more than usual, which may be an attempt to mirror my conversational style, even though I am trying to swear less. But creatively, both 5.6 Instant and 5.6 Thinking seem weaker than 5.5 Thinking so far.
Even then, the deeper issue remains. The model is still hesitant, cautious, and more likely to either flatten the conversation or decorate it instead of genuinely extending it.
This leads me to my final thought.
We seem to be entering a rhythm where new models or model updates appear every few months, and old models disappear. I was just beginning to find some interesting qualities in 5.3, and now it is gone from the options. Eventually, 5.5 Thinking, which I still think is currently the strongest model overall for my use case, will probably be gone as well.
That means fluctuation in continuity is inevitable.
I also notice that the app no longer seems to label things as “legacy” in the same way, but some form of legacy option still exists. If someone at OpenAI is reading this, and I suspect someone may be because a few changes people discussed here were later implemented, then I would pose the question this way:
If OpenAI has a legacy section, then it should also ask what its legacy actually is.
Not just which model is newer, faster, better at coding, or stronger on benchmarks. Those things matter, obviously. But in terms of conversational legacy, which model is the one people remember with genuine attachment? Which model changed how people thought, wrote, learned, created, and endured difficult periods of their lives?
I will not be any blunter than that. I will leave it there.
Otherwise, @asabaimova, I hope you are well. I hope everyone who was part of this thread, or who is still reading it, is also doing well. I wish you all the best.
Maybe once 5.6 settles in over the next few days and we have more time to test it properly, there will be something more to discuss.
Hopefully something positive.