Problem Solutions
Why GPT-6 Feels Worse Than GPT-5.6 Sol at Memory and Long Chats — and What Actually Helps
Since the GPT-6 rollout, a growing number of ChatGPT users say the new model is worse than GPT-5.6 Sol at the thing long chats need most: remembering. A detailed comparison thread on the OpenAI forum from October 20, still active on October 23, and a follow-up bug report from October 27 describe the same pattern. These are user reports, not OpenAI-confirmed regressions — but they are specific enough to act on.
What users report
GPT-5.6 Sol, the comparison says, picks up implied intent quickly, follows conversational rhythm, preserves established tone and answers the actual point without unnecessary framing. GPT-6, by contrast, feels like it is standing outside the conversation and describing it: it restates things that are already obvious, adds unnecessary explanation, flattens distinctive language into generic prose and turns natural exchanges into something padded and overmanaged. The gap grows in long conversations: context and continuity degrade sooner, established details become less reliable, responses become more generic, and the useful lifespan of a chat feels dramatically shorter. Memory feels less dependable too: information that should stay part of the conversation is more likely to be ignored, flattened or replaced with a generic approximation.
Other threads add more: responses sometimes shorter than expected, detailed instructions ignored or only partly implemented, explicitly requested events or dialogue omitted, long responses ending early. One user reports six consecutive days of deterioration from October 7–12; another says OpenAI support redirected them to basic troubleshooting — Memory, Custom Instructions and Project settings — instead of addressing the regressions. Users also ask OpenAI to keep GPT-5.6 Sol available as a selectable model and not to substitute GPT-6 silently when a user deliberately chose 5.6.
What this does not prove
Forum posts cannot confirm a technical cause. The problem could sit in the model, in routing, or in a shared backend component — one tester found the same disappearing-response behavior on GPT-5.5, which points away from a single-model explanation. It also does not establish that every user is affected: several reporters stress the failures are intermittent. OpenAI has not confirmed a GPT-6 memory regression in these threads, so treat the pattern as reported, not established.
What actually helps
These are workarounds users say help, not guaranteed fixes. Switch to GPT-5.6 Sol while it is selectable, and check which model actually answered — do not assume your choice stuck. Start a fresh chat when a long one degrades; it does not always fix the problem, but it is the first step. Keep a running summary: paste the key facts, decisions and constraints at the top of a new chat, or keep them in a document you can re-paste. Use Projects and custom instructions that emphasize completeness and continuity, and repeat critical instructions inside the chat rather than relying on settings alone. Break long work into smaller chats per topic, and save important outputs outside ChatGPT so a degraded chat cannot take them with it.