You send a long prompt or upload a PDF, and ChatGPT stops halfway with “Error in message stream” or “Unknown error occurred”. The chat is fine; the connection that carries the answer is not. This guide explains why that happens and how to fix it on desktop, mobile, and file uploads.
Quick answer
“Error in message stream” means the reply stopped arriving before it finished.
To fix it:
- (1) check status.openai.com,
- (2) click Regenerate,
- (3) turn off your VPN,
- (4) open the chat in a private window,
- (5) clear cache and cookies.
If it only fails in one chat or Project, start a new chat.
What “Error in Message Stream” means
ChatGPT does not send a finished answer in one piece. It streams the reply to your screen a few words at a time over an open connection. If that connection breaks for any reason, even for a moment, the reply stops and an error appears.
The exact wording “Error in message stream” is what users call it, and it is not a formal error name. OpenAI’s help documentation describes the same family of failures under messages such as “A network error occurred”, “An error occurred while connecting to the websocket”, “There was an error generating a response” and “Something went wrong”. When you see an error, note the exact text, because the best fix depends on it.
None of these mean your account or chat is damaged. They mean one request did not complete.
Common errors and what to do
| Message | Likely cause | Try this first |
|---|---|---|
| Something went wrong / Unknown error | Temporary server issue, or a browser or session problem | Refresh, log out and in, try another browser, check the status page |
| Network error / WebSocket error | Weak connection, VPN, proxy, or firewall | Turn off VPN or proxy, switch network, try a private window |
| Suspicious activity detected | Traffic from a flagged VPN or shared IP, or a shared account | Disconnect the VPN, restart the browser, stop sharing the login |
| Problem preparing your chat | An extension is blocking scripts on the page | Disable ad blockers and script tools, or use incognito |
| Error generating a response | The reply failed partway | Click Regenerate once before changing anything else |
| Unknown error on file upload | Large, scanned, corrupted, or unsupported file | See the upload section below |
The question to ask is whether the problem is on your side (network, browser, file) or on OpenAI’s side (outage, overload). The steps below start with the cheapest checks.
Step-by-step fix
-
Check the status page
Open status.openai.com. If it reports degraded performance or an outage, wait. Nothing you change locally will help.
-
Regenerate, then refresh
A one-off failure often clears with Regenerate. If not, reload the page (
Ctrl+RorCmd+R). In the app, force-close it and reopen. -
Test your connection
Switch between Wi-Fi and mobile data. Turn off any VPN or proxy, since these are the most common cause of network and WebSocket errors.
-
Use a private window
If the error disappears in incognito, the cause is an extension or stored browser data. Ad blockers, script injectors and AI writing extensions are frequent offenders.
-
Clear cache and cookies
Open your browser’s privacy settings, clear cached files and cookies for chatgpt.com, then sign in again. This also renews a stale session token.
-
Shorten the request
Very long, multi-part prompts raise the chance of a timeout. Split the task into smaller messages and send the next one only after the previous reply finishes.
-
Try another browser or device
If it works elsewhere, the problem is local to the first browser. Update it or reset its settings.
-
Contact support
If the error persists on several browsers and networks, go to help.openai.com and report it. Include a screenshot, your browser and version, the time it happened, and whether it occurs with plain prompts or only with files.
“Unknown error occurred” when uploading files
This version appears while a file is being uploaded or read, before ChatGPT starts answering. The same file can fail once and work a minute later, which usually points to processing or connection trouble rather than your plan.
Why uploads fail more often
- Parsing: ChatGPT has to extract text first. Scanned, image-only, encrypted or damaged PDFs can fail at this stage.
- Timeouts: Large files take longer to process and may exceed time or token limits.
- Browser memory: Handling a big file locally can fail on low-memory devices or crowded tabs.
What to try
- Re-upload after a refresh, or wait a few minutes and try again.
- Rename the file with plain letters, numbers, and hyphens.
- Run OCR on scanned PDFs so they contain real text.
- Split large documents into parts, or export only the pages you need.
- Recreate a stubborn PDF with Print, then Save as PDF, which strips hidden formatting.
- Wait until the upload finishes before you send your prompt.
File-size, token, and image limits change over time and differ by plan, so confirm the current numbers in OpenAI’s Help Center. As a habit, keeping files well under the stated limits makes uploads more reliable.
Fixing it in Projects and long chats
Some people see the error only inside one Project or one very long conversation, while a fresh chat works fine. OpenAI has no official fix for this case, but these community workarounds are low risk and delete nothing:
- Start a new chat inside the same Project and paste in only what you need.
- Move the failing chat out of the Project, send a message, then move it back.
- If a whole Project keeps failing, create a new one and re-add its files.
Treat these as workarounds, not guarantees. Check the status page and release notes first, because OpenAI has fixed streaming bugs on its side before.
Work and school networks
Streaming needs a connection that stays open for the length of the reply. Corporate firewalls, filtered DNS, proxies and VPNs sometimes cut long-lived connections early, which ends the stream mid-answer. If ChatGPT works on your phone’s hotspot but fails on the office network, the network is the cause. Ask your IT team to allow chatgpt.com and its WebSocket traffic, or use a different network.
For developers using the API
- Log the status code and how long the stream lasted, so you can tell a dropped connection from a bad request.
- Retry dropped streams with exponential backoff and a retry limit.
- Handle partial output: keep what arrived and decide whether to resume or restart.
- Check that no proxy or load balancer in your path buffers or times out streaming responses.
Why the stream breaks
- Interrupted connection: Even a short drop ends the stream, and the interface reports it as an error.
- Rate limits and load: When demand spikes, requests can be reset mid-reply.
- Slow file handling: If parsing takes too long on either end, the request times out.
- Model or rollout changes: During updates, queued requests can be dropped.
How to prevent it
- Stay on a stable network and avoid VPNs while chatting.
- Keep prompts focused. Several smaller requests are safer than one huge one.
- Wait for each reply to finish before sending the next message.
- Copy important output as soon as you have it, so a failure does not cost you the work.
- Reload long-idle chats before continuing, since sessions can expire quietly.
- Keep your browser updated. Current Chrome and Edge builds are well supported.
- Developers: if you use the API, add retries with exponential backoff to handle dropped streams.
Frequently asked questions
Q1. What does “Error in message stream” mean?
Ans. The connection that delivers ChatGPT’s reply was interrupted. Causes include a weak network, an expired session, a VPN, or a temporary server reset.
Q2. How do I fix it quickly?
Ans. Check the status page, click Regenerate, refresh, and try a private window. If it still fails, clear the cache and cookies or switch browsers.
Q3. Why does it happen when I upload a PDF?
Ans. Usually the file is too large, scanned without text, damaged, or oddly formatted. Use a text-based PDF, rename it, and split it if it is long.
Q4. Is the problem on my side or OpenAI’s?
Ans. It can be either. If the error follows you across browsers and networks, or the status page shows an incident, it is likely on OpenAI’s side. If it disappears in another browser or without a VPN, it is local.


