ChatGPT connection interrupted: the answer often keeps going, how to get it back, and what 19 forum threads show

What does ChatGPT's connection interrupted message mean?

OpenAI's help center doesn't list "Connection interrupted" anywhere, as of October 2026, so its meaning comes from one support reply. The OpenAI Support account wrote it on the OpenAI community forum on 16 October 2026, under a developer's log of about 187,000 connector calls: the banner "doesn't necessarily mean the MCP run stopped." In plain terms, ChatGPT is trying to reattach to the same reply, and the model may keep working while your screen waits.

We searched the forum for "Connection interrupted" on 18 October 2026 and read every thread that quotes the banner: 18 from the search, plus a July thread titled with the message that the search skipped. Each thread counts once, by what the first post quoting the banner says was running. The first one we found opened on 29 June 2026. That July thread, a single post, has 7,300 views. Its author was coding with ChatGPT and asked why it "needs me to be connected while thinking, while im totally connected".

ChatGPT translates the banner along with the rest of its interface. Traditional Chinese shows "連線已中斷。正在等待完整回覆" (quoted in a GitHub report on 8 October 2026), and Dutch shows "Verbinding onderbroken". Some users paraphrase it as "waiting for the full answer" or "waiting for the full response". Reports place it on chatgpt.com in Chrome on Windows, macOS and iPad, in the iPhone app and in the Mac app.

How do you fix ChatGPT connection interrupted?

Don't press anything for the first few minutes. A reload doesn't cancel a reply that's running on OpenAI's side, but it can lose text you haven't sent yet. Work through these in order.

1. Copy your unsent text and any partial answer. Nothing after this step can then cost you the prompt.
2. Wait at least a minute, longer for a Thinking model. OpenAI's troubleshooting page says to give a hanging reply 30 to 60 seconds before you do anything. A long reasoning run can need more.
3. Open the same chat on a second screen. Use a new tab, or the ChatGPT app on your phone. If the reply is growing or finished there, the first screen only lost its stream.
4. Reload the original tab. On the web this often brings in the finished reply, or the progress so far.
5. If the banner stays and nothing grows for 10 minutes, press Stop, then Regenerate. OpenAI's troubleshooting page gives that same order for a reply that hangs.
6. Check status.openai.com or Whizi's free Is ChatGPT down? page. Read the Conversations and ChatGPT Work rows, not just the banner at the top.
7. Shorten the job. Pick a lower reasoning level, or ask for the output in numbered parts, so each reply ends well before the 25 minute mark covered below.
8. Rule out your connection. Turn off a VPN or proxy, clear the site's cookies and cache, and try your phone on mobile data. Then send OpenAI support the chat link, the model and the time in UTC.

Step 3 matters most. In the 30 July thread on a custom connector, the first poster wrote that a refresh "reveals the finished answer". In the 12 October follow up, that developer's own server kept receiving tool calls from the chat "for minutes after the banner appears". The reply was still being written. Only the screen had stopped.

Is the answer lost when ChatGPT says connection interrupted?

Usually not: the reply is often still running, or already saved, when the banner shows. A GitHub report from 8 October 2026 is the clearest case. A ChatGPT Pro user ran a long website audit, and the work finished and saved its final files around 11:37 Amsterdam time. The chat still showed the turn as active, with the reconnect message, "3h42m" in. Refreshing didn't correct the screen. The work had still finished.

The answer is lost in a minority of cases, and what you see after a reload tells you which case you're in. The table below lists every report we read that quotes the banner and says what happened next, as of 18 October 2026. Model names are as the posters gave them.

Read it in three groups. If a reload shows the full reply, you're done. If it shows partial progress and the banner stays, the reply is still running, so wait or check from your phone. If you get "Resume stream unavailable", "Could not load this ChatGPT conversation" or a blank turn, ChatGPT couldn't reconnect to that reply. Paste your prompt into a new chat and run it again in smaller parts. A chat that won't load at all is a different error, and ChatGPT response not loading covers it.

Why does ChatGPT get interrupted on long Thinking answers?

Long Thinking runs give the stream more minutes to break, and since August 2026 some of them have stopped near the 25 minute mark. The 209 post forum thread opened on 22 August 2026 logs ordinary chat turns on GPT-5.6 Thinking at High ending at 1,541 to 1,564 seconds. That's 25 to 26 minutes. The same poster had kept an earlier chat turn that ran 102 minutes and 26 seconds. A separate September thread on the iPad and phone apps says the cut came "exactly at 25 minutes".

One trace in that thread shows the banner arriving after the work had already died. On 19 September a poster ran GPT-5.6 Sol at maximum effort. The last useful tool result came at 24 minutes 34 seconds. Nothing followed but heartbeats. At about 45 minutes the client asked to resume the stream and got a 404 reading "Resume stream unavailable". The screen then showed "Connection interrupted", and no answer came. The poster says the trace shows timing, not cause.

So the 25 minute pattern changes the advice. If your request needs more than about 20 minutes of reasoning, split it. Ask for part one, check it, then ask for part two. OpenAI Support hasn't confirmed any time cap, and the thread's posters say the stop point moved between accounts and weeks. For long tasks, a new chat with a summary of the old one also helps. The ChatGPT conversation too long fix shows how to carry the context over.

Does connection interrupted mean ChatGPT is down?

Connection interrupted rarely means ChatGPT is down. It's OpenAI's side when status.openai.com shows an open incident on Conversations or ChatGPT Work. It's probably your setup when the same chat works on your phone over mobile data. Neither test is perfect for this banner.

The status page rarely names it. We read every incident on status.openai.com's history on 18 October 2026: 37 incidents from 22 September to 16 October, and none used "interrupted" or "connection" in its title. OpenAI writes them as errors instead. On 8 October it posted "Some APAC users are seeing errors in ChatGPT Work, conversations, and GPTs.", which ran 94 minutes. On 4 August an incident on Conversations ran 162 minutes and said: "Some users may experience failed or interrupted responses, and some conversation turns may not be saved." That's the case where waiting doesn't help.

The device test has a twist too. In the connector thread the web client dropped the stream while the mobile app, on the same server and connector, never did. If one surface fails and another works, the reply is fine and the failing surface is the problem. "Network connection lost. Attempting to reconnect" is a different message about your own link. It belongs to the ChatGPT network error guide, where a VPN or office firewall is the usual cause. A general "Something went wrong" has its own steps in the something went wrong fix.

What if long answers keep getting interrupted?

Finish the urgent part somewhere else, then come back to ChatGPT in smaller pieces. What to use when ChatGPT is down lists the options and how to move your context across.

Whizi, which publishes this page, runs each reply on its own server, not in your tab. Closing the tab or losing the connection doesn't end the answer, and a reopened tab or another device picks the stream back up. If you press Stop, Whizi saves the partial answer and refunds a turn it never delivered. Where it loses: its GPT answers still come from OpenAI's models, so an OpenAI incident reaches them too. Whizi isn't immune either: a stream that breaks there shows "The model stream was interrupted." ChatGPT Luna and Claude Haiku both come with the $15.99 Starter plan.