Problem Solutions
Why ChatGPT Keeps Breaking Mid-Task — and How to Stop Losing Your Work
You are halfway through a long job in ChatGPT when the response stops. "Streaming interrupted. Waiting for the full message." A few minutes later: "Error in request. Please try again." Or the chat itself gives up, declaring "This chat has become too long" after only a few messages. Since mid-October, OpenAI's own community forum has filled with reports of exactly this pattern — and the common thread is not one bug, but three different ways a job can die.
The three ways work dies
First, the stream interrupts and never resumes. Users describe responses that cut off after about two minutes of thinking, sometimes with a "Connection interrupted" banner, sometimes with nothing at all. One user who checked the network logs found the session ending on a token limit; another found that the turn sometimes keeps running server-side for minutes after the banner appears, and refreshing shows the progress made so far — but the banner stays until the turn ends.
Second, capacity and rate errors kill the session. After about five or six minutes of work, some users get "Error in request" every time; others hit "Could not load this ChatGPT conversation" across every open tab at once, with retry doing nothing. One thread from October 17 describes a multi-tab lockout: one tab timing out flagged the session token and killed access across all open ChatGPT tabs simultaneously. Users on plans from Plus to the $100 tier report the same failures, and several note afternoons in Europe — when US users come online — are the worst.
Third, the chat itself runs out of room. "This chat has become too long" now appears after only a few messages for some users, forcing them to start new chats repeatedly — one user needed eight new chats in three days. When a long session does crash, the reload can come back with the context gone, so the job's state — what was decided, what was already done — is lost with it.
How bad has it got? One user who posted on October 19 put it bluntly: two months ago ChatGPT could work for 45 to 60 minutes; a month ago that was down to 22 to 25 minutes; now, they wrote, "I have it EVERY 4-6 minutes." OpenAI staff have acknowledged in community discussion that the issue is being investigated, but users report no clear explanation, resolution timeline or official workarounds so far. This article is based on those user reports, not on an OpenAI announcement — the company has not published a root cause.
What actually helps
None of this fixes a server-side capacity problem. What it does is limit how much you lose when the session dies, and reduce the chances of triggering the failure yourself.
Work in small chunks with checkpoints. Break a long job into steps that each finish in a couple of minutes, and after each step ask ChatGPT to summarise what is done and what comes next. Short turns are less likely to be cut off than one long thinking run, and a checkpoint gives the next session something to resume from.
Keep a handover record outside the chat. The fatal flaw in all three failure modes is that the job's state lives only inside the session. Keep a short note — goal, decisions made, files changed, next step — in your own document, and paste it into a fresh chat when one dies. A dead session then costs you a paste, not the whole job.
Use one chat per workstream, and one tab for heavy work. Very long chats hit the length flag sooner, and multi-tab sessions can lock each other out. Start a fresh chat for each distinct task instead of growing one mega-chat, and avoid running heavy jobs in many tabs at once.
When a stream breaks, wait before resending. The turn may still be running server-side. Refresh once to see whether progress appeared, and if the response truly stopped, ask it to continue from the last checkpoint rather than restarting the task from scratch.
Save outputs to files as you go. Code, drafts and results should land in your own files after every step, not sit only in the chat transcript. If a conversation becomes unloadable, the work survives even if the chat does not.
If the failures persist across days, check OpenAI's status page and the community forum before assuming your account is at fault: the pattern in October has been service-wide, not individual. And for workloads that must not die mid-task, the API or Codex — where limits are explicit and billed, rather than silently enforced in the chat interface — may be the more reliable route.