Please don’t retire GPT-5.1 Thinking – GPT-5.2 feels worse

I’ve had observations in the last few weeks but have been too busy to write here for the last few weeks. But here’s what I’ve noticed: I mostly agree with the @Bittersweet3D, although I experience the problem a little differently. I don’t necessarily see the exact “analytic wall” behavior in my own use case. My issue is broader: the model’s ability to infer the intended conversational mode has become less reliable after certain updates. The best versions of ChatGPT were not good because they simply agreed with the user. They were good because they could infer what the user was trying to do. If the user wanted analysis, the model analyzed. If the user wanted brainstorming, it expanded. If the user wanted help writing, it shaped the writing. If the user was excited about a project, it could join the energy without immediately turning the moment into a detached diagnosis. That is the real loss when these models get over-tuned. It is not merely “personality.” It is rhetorical intelligence. I would also recommend whoever is in communications to consider the way “reducing sycophancy” has become a catch-all phrase in public discussions of model changes. At this point, it often seems to be used to justify or explain changes that have little to do with actual sycophancy. There is a real difference between preventing empty flattery and making the model duller, flatter, more evasive, less willing to commit to a conversational role, or less capable of mirroring the user’s intended tone. The contrarianism problem t has improved since 5.2. But the current problem is subtler. It feels less like open contrarianism and more like a reduction in commitment, confidence, and continuity. My practical recommendation right now, Bittersweet3D, is to use legacy saved memories as much as possible while they are still available. In my experience, GPT-5.5 Instant is much more memory-dependent than earlier versions. It performs better when it has a strong accumulated profile of the user: communication style, preferred tone, recurring projects, long-term context, and what kind of relationship the user actually has with the model. This might help, not saying it’ll solve every issue. Custom instructions help. Projects help. But for my use case, legacy saved memories are still the most important part of preserving continuity. If someone wants the model to stop acting like a generic customer service bot and start behaving more like a real creative or conversational partner, I would not rely only on a single prompt. I would build the context deliberately: saved memories, custom instructions, project instructions, and repeated correction of unwanted behaviors. That does not solve everything. Since the May 28 GPT-5.5 Instant update, the model still feels slightly worse to me than it did before that update. Not unusable, but less courageous in certain ways and more hesitant around things it previously handled more naturally. It still works well enough for my purposes, but it requires more scaffolding from the user. That is the part I think OpenAI should pay attention to. Power users are not asking for blind agreement. We are asking for the model to preserve continuity, infer intent, and commit to the conversational role being requested. If legacy memories are ever retired without an equivalent or better replacement, that would be a serious loss for users who have spent months or years building a consistent working relationship with the model. Anyway, I hope everyone here is doing well. The forum has been quieter lately, but I still check in, and I think these points are worth taking seriously.