ChatGPT error in message stream: what OpenAI does and does not say

What the message means, and who says so

OpenAI’s two help pages on errors, the troubleshooting article and the network recommendations, do not contain the words “message stream” in the text we read. AI Toolbox, a blog, says the same in terms: “OpenAI has no Help Center article for this exact message.” What the blogs add is an explanation. Axentix says ChatGPT streams its reply instead of waiting until it is complete, and that “Error in message stream”, sometimes shown as “Error in body stream”, is that stream failing. AI Toolbox puts it as the stream of text from OpenAI’s servers to your app breaking before the reply finished. These are explanations by third parties, consistent with each other, and we cannot check them against an OpenAI page.

What OpenAI’s page says about the errors it does list

The troubleshooting article opens with the causes: “ChatGPT may occasionally encounter technical issues due to factors like network configuration, browser extensions, or transient server-side problems.” For a general error, which it says “may indicate a temporary server issue or a problem with your local setup”, its steps are to refresh the page or start a new chat, check status.openai.com for known outages, clear the browser cache and cookies, use an incognito window, disable browser extensions, especially privacy or security tools, turn off VPNs or secure DNS tools, and switch to a different network.

For “A network error occurred.” and “An error occurred while connecting to the websocket.”, which the page says suggest that your device is unable to establish a reliable connection to OpenAI’s servers, it adds turning off Web Protect or other security filters and trying another browser or device. For “There was a problem preparing your chat.” it points to an extension or URL blocker; for “There was an error generating a response.” it says it may be a one-time error and to click Regenerate; and for a chat stuck on Thinking, or a slow one, it adds starting a new chat if the conversation is long. The existing page on why ChatGPT is slow covers that last sequence.

On a company or school network

OpenAI’s network article is written for IT departments and for people on a company Wi-Fi. It says: “Some ChatGPT and Codex features use secure WebSocket connections in addition to standard HTTPS requests. If your organization blocks WebSocket traffic by default, some features may stall, disconnect, or fail to stream updates correctly.” It asks for WebSocket traffic over TCP port 443 and a proxy, firewall or secure web gateway that permits the “Upgrade: websocket” handshake to ws.chatgpt.com for conversation updates and notifications. For networks that inspect TLS or enforce a proxy, it asks that these controls do not block, rewrite or prematurely close the WebSocket handshake or connection; if sessions connect but later stall or disconnect, check the idle timeout and the maximum message-size policies for WebSocket traffic. It lists security gateways it knows are used to filter web traffic, such as McAfee Web Gateway, Cisco Umbrella and Zscaler Internet Access, and says that commonly the software has to stop blocking ChatGPT or explicitly allowlist OpenAI’s domains.

If the message appears only on one network, that is the page to give the administrator. A message that appears everywhere is a different case: check the status page first. For a similar error in another product, the page on Roblox error 524 follows the same order, from the account to the network.

What the blogs add

AI Toolbox, dated 8 October 2026, is a blog of a Chrome extension for ChatGPT, Gemini, Claude and Grok. It says the message is usually server-side, during an incident, on a very long reasoning run, or in one broken chat or Project; that you should click Retry first; that if one chat fails every time you should start a new chat and carry over a short summary; and that the only status incident named after the message was on 11 June 2025 and was fixed by rolling back a change. Merca20, a news site dated 10 September 2026, reports surges on Downdetector. Axentix, dated 18 April 2026, mentions an incident in November 2025. None of this is on the OpenAI pages we read, and we did not check the status history ourselves.

OpenAI Developer Community threads about the message appear first in the search results, but the pages we could read were empty shells, so we do not report what they say. OpenAI’s pages and AI Toolbox give the same first steps: check the status page and start a new chat.