Another related issue I think OpenAI needs to address is the experience with very long conversations in normal ChatGPT.
I use long-running chats because keeping the same conversation preserves important context. But eventually they become genuinely annoying to use. The browser can become increasingly laggy as the conversation grows, scrolling and typing get slower, and I have personally seen a long ChatGPT tab take around 1 GB of RAM.
Other users have reported similar problems with long conversations freezing, hanging, or even crashing the web client, so this does not seem unique to my machine.
I also don’t understand why the solution should effectively be “use Work” when Work and Codex consume a separate shared agentic allowance.
Work is actually much better suited to long, complicated tasks and doesn’t seem to suffer from the exact same long-chat problem in the same way, but now using that workflow consumes a limited resource. That is frustrating when similar long-running work could previously just be done inside normal Chat as part of the subscription.
I think there are two separate problems OpenAI should fix:
-
Normal Chat should handle very long conversations better. The browser should not need to keep an enormous rendered conversation in memory at once. Older messages could be virtualized, paginated, or unloaded from the page while still remaining part of the conversation.
-
Users shouldn’t be pushed toward a metered Work mode just because normal Chat becomes unusable in long threads. Work should provide additional agent capabilities, not effectively become the paid workaround for Chat’s long-conversation performance problems.
And if there is a practical conversation-length limit in normal Chat, I would like OpenAI to reconsider how restrictive it needs to be. Codex and Work already demonstrate that extremely long-running workflows can be handled through context management instead of simply making the user abandon a conversation.
Long conversations are especially valuable for coding, research, and projects where starting a new chat means losing a large amount of accumulated context. Improving this would make ChatGPT substantially more usable without requiring more model intelligence at all—it is largely a product/client architecture problem.