Four different errors get lumped together here and they have different causes. “Something went wrong” is usually Anthropic’s end and retrying fixes it. “An unknown network error” is usually yours, and a VPN is the most common reason. Error 529 means overloaded, so wait rather than retrying hard. Before doing anything else, check status.claude.com, because if it is an outage nothing you do locally will help.
Claude fails to respond and you get an error. Which error you got matters, because the causes are not related to each other and neither are the fixes.
The messages are also unhelpfully vague, and two of them sound like they mean the same thing while pointing at opposite ends of the connection.
Thirty seconds at status.claude.com first. If something is down there, stop reading and wait, because none of this is about your setup.
Which error are you seeing?
“Something went wrong. Try sending your message again.”
Almost always Anthropic’s end. Retry the message. If it persists past a few attempts, check the status page, then try a new chat.
“An unknown network error has occurred”
Usually your end. VPN first, then extensions, then your network. Fixes 2 to 4.
Error 529, or a message about being overloaded
Anthropic is at capacity. Waiting is the fix. Hammering retry makes it worse.
Error 503, or the page will not load at all
Service unavailable. Check status.claude.com. If it is green there, it is your connection or your network blocking access.
Fix 1: check the status page first
status.claude.com shows whether Claude is actually up. It is the fastest way to find out whether you are troubleshooting your own setup for no reason.
Worth knowing that the web app, the API and Claude Code can be affected separately, so a green light on one does not clear the others. Read which component the incident is against.
Fix 2: turn off your VPN
If you are on a VPN and getting network errors, turn it off and retry. This resolves a large share of “unknown network error” reports on its own.
Sustained requests are the first thing an unstable tunnel drops, and Claude’s responses stream over a connection held open while it writes, which makes them more fragile than an ordinary page load. If you need the VPN on, try a different server, ideally one geographically closer to you.
Fix 3: rule out extensions
Open claude.ai in a private window and send one message. Extensions are usually disabled there.
If it works in private and fails normally, an extension is the cause. Ad blockers and privacy extensions are the usual suspects because they interfere with exactly the kind of persistent connection Claude uses. Re-enable them one at a time to find which.
Fix 4: clear the site data for claude.ai
Claude.ai keeps a copy of its own code in your browser. When that copy goes stale you keep running an old version, faults included.
Do not clear your whole browser cache. Clear this one site:
- Open claude.ai
- Click the icon to the left of the address bar, the padlock or the sliders
- Choose Cookies and site data, then Manage on-device site data
- Delete everything listed
- Hard refresh with
Ctrl + Shift + R, orCmd + Shift + Ron Mac
You will be signed out. Your conversations live on Anthropic’s servers, not in your browser, so nothing is lost.
Error 529 and being overloaded
529 means the service is at capacity. It is not about your account, your plan or your request.
The unhelpful truth is that waiting is the fix. Retrying immediately and repeatedly adds load to the thing that is already overloaded, and if you are using the API it can push you into rate limiting on top of the original problem. Back off, wait, and space out retries with increasing gaps rather than firing them one after another.
Peak hours make this considerably more likely, so if the work can wait an hour it often just resolves.
When a long conversation is the trigger
If errors started appearing in one particular chat and everything else works, the chat is the problem rather than your connection.
Very long conversations with a lot of attached material get heavy, and a heavy request is more likely to time out. Start a new chat and carry the context over deliberately: ask Claude to summarise where things got to, then paste that as your first message. If the new chat works immediately, that was it.
Work and school networks
If Claude fails at work and works at home, it is the network, not Claude.
Test by tethering to your phone for one message. If that works, ask whoever runs your network to allow claude.ai and anthropic.com. Mention that Claude streams responses over a persistent connection, because proxies configured to buffer or inspect traffic often break that specifically while leaving ordinary pages working fine, which makes the problem look intermittent rather than blocked.
Why the messages are so vague
The app can tell that something failed and often cannot tell what. A dropped connection, a proxy closing a stream, a timeout and an overloaded server all arrive as the same event from where the browser is standing: the response stopped.
So it reports something generic. That is why the same message covers faults at opposite ends of the connection, and why working out which one you have is most of the job.
Written by the team at Lead Source, who spend their days on lead attribution rather than error codes, but keep a status page bookmarked like everyone else.