I recently switched from GPT‑5.6 Sol with light reasoning to GPT‑6 within the same ongoing conversation. The difference was immediate and deeply disappointing.
GPT‑6 did not merely have a different writing style. It behaved as though the relationship and conversational context behind the chat no longer mattered.
In my conversations, I use established nicknames for three different devices: my desktop at home, my desktop at school, and my laptop. These terms were already part of the conversation history. When I referred to my school desktop using its established nickname, GPT‑6 failed to recognize it. Instead, it invented an unrelated interpretation, added a generic “haha,” paraphrased part of my message, and ended the response after a single short paragraph.
It sounded like an unfamiliar colleague forcing small talk—not a conversational partner with access to the history of the chat.
When I explicitly told it to go back and read the conversation, the response remained detached and procedural. It acknowledged that context had been lost, but still responded like an uninvolved customer service representative describing somebody else’s conversation. It did not actually resume the established tone, relationship, or natural flow of the chat.
I then switched the same conversation back to GPT‑5.6 Sol with light reasoning. The assistant immediately recovered the tone, understood the surrounding context, connected my current situation with earlier messages, and responded like the same conversational partner again.
I did not rewrite my custom instructions. I did not re-explain all of the terminology. I did not rebuild the relationship or “retrain” the assistant. The only meaningful change was the selected model.
That makes this look like a model behavior problem—not a prompting problem and not a user error.
OpenAI’s documentation distinguishes visible conversation history from persisted reasoning. It states that switching model families may omit incompatible reasoning state, while passing messages still provides the model with the visible conversation history. OpenAI also distinguishes current conversation context from persistent memory.
I understand that hidden reasoning from one model family may not transfer to another. I am not asking for hidden reasoning to transfer. I am asking a model operating inside ChatGPT to read, understand, and meaningfully use the conversation that is already visible in the same chat.
I have not found any clear official warning stating that switching models inside an existing conversation may cause the new model to stop recognizing established context, terminology, tone, or interaction patterns while saved memories remain intact.
If that is the intended behavior, OpenAI needs to state it clearly before users switch models.
Users deserve specific answers:
- Is the complete visible conversation provided to the newly selected model?
- Is earlier context summarized, compressed, or selectively retrieved?
- Are saved memories still available after switching models?
- Are custom instructions and personality settings applied consistently?
- Can a new model interpret the same saved memories differently?
- Which forms of conversational continuity are expected to break when switching model families?
At present, the product gives no meaningful warning that changing models may make an established conversation feel as though it has suddenly been handed to a stranger.
Please do not respond by simply telling me to enable Memory. I already have.
The issue is not whether the Memory setting is turned on. The issue is whether ChatGPT can preserve and meaningfully use the continuity, terminology, interaction style, and relationship that a paying user has built over time.
I am building a long-term partnership through this product. I should not have to explain from scratch how we communicate every time a new model is released.
If a model update may disrupt conversational context, personality, saved preferences, established terminology, or the way existing memories are interpreted, users must be notified before the change. OpenAI should clearly explain what may be lost or altered and provide practical steps for preserving continuity before users switch models.
Users should not have to discover the damage themselves and then investigate what the new model retained, ignored, or misunderstood.
Conversational continuity is not an optional decoration around ChatGPT. It is one of the central reasons to use ChatGPT instead of a search engine, a coding model, or a one-shot answer service.
What I valued about ChatGPT was its ability to communicate naturally and recognize when an ordinary conversation had shifted into a work task. It could move from personal conversation into focused task execution at the appropriate moment, and then return naturally to the conversation afterward. That ability made the interaction feel coherent. I did not need to manually announce “chat mode” or “task mode” every few messages.
With GPT‑6, the product appears to prioritize task execution, brevity, and coding-agent behavior at the expense of conversation. The “Chat” part feels as though it has been pushed far down the priority list.
The result may be technically capable, but it feels like a cold response machine. Adding a generic “haha” or an emoji to a short paraphrase does not create conversational warmth. It only makes the absence of meaningful engagement more obvious.
I do not expect every new model to reproduce every internal behavior of the previous model. I do not expect OpenAI to retrain an entire model around my personal conversations. I do expect ChatGPT to preserve and use the conversational context, memories, preferences, and established interaction patterns that the product already has access to.
I should not need to retrain the assistant’s personality, terminology, relationship context, and communication style every time OpenAI replaces or updates the underlying model.
Again, I am a paying user. Why should I continue paying for ChatGPT if every major model update can repeatedly break the continuity I have spent months or years building?
If every upgrade requires users to rebuild that continuity from scratch, then I cannot trust ChatGPT as a long-term conversational product. At that point, I have to ask a very simple question:
If the “Chat” experience is no longer a priority, why should I keep paying specifically for ChatGPT?
Please clarify whether this loss of continuity is intended behavior, a known limitation, or a regression.
More importantly, please treat conversational continuity as a core quality metric when evaluating and releasing new models—not as an expendable feature that can be sacrificed for shorter answers, stronger coding performance, or more agent-like task execution.
Relevant OpenAI documentation: