2 min read AI-generated

Claude Down Worldwide: Three Hours of 529 Errors

Copy article as Markdown

On July 29, Claude was disrupted worldwide for about three hours. Instead of answers, users got a wall of '529 Overloaded'. It hit claude.ai, the API, and Claude Code — so every tool built on Claude went down with it.

Featured image for "Claude Down Worldwide: Three Hours of 529 Errors"

You know the moment: the cursor blinks, you wait for a reply — and all you get is an error. On July 29, a lot of people had that moment at the same time. Claude was down worldwide for roughly three hours.

What happened

Anthropic reported elevated error rates across multiple models starting at 19:49 UTC. Anyone using Claude ran into the same message over and over: ‘API Error: 529 Overloaded’. A 529, at its core, means the servers can’t keep up with the load. At 20:33 UTC Anthropic said it had identified the issue and was working on a fix — without naming the cause. By around 22:36 UTC the service was fully restored.

This wasn’t just the chat interface. The outage hit claude.ai, the API, and Claude Code all at once. And because countless tools quietly reach for Claude through the API in the background, they stalled too. The top comment under the BleepingComputer article said it all: ‘Ughh how will I code now.‘

Why it’s more than a hiccup

A three-hour outage isn’t a catastrophe by itself. But it shows how deeply Claude now sits inside real workflows. Two years ago a chatbot outage was an annoyance. Today it freezes half the toolbox for a lot of people — the coding agent, the automated scripts, the Cowork sessions running in the cloud.

It’s not the first outage this year, either. There were similar disruptions in March and April. The pattern behind them is usually the same: capacity. When demand grows faster than infrastructure, ‘529 Overloaded’ becomes the symptom of your own success.

My take

I always walk away from days like this with the same lesson: don’t build a critical process that only works with a single provider. A fallback model, a saved local state, a plan B for when things break — it sounds like extra work, but it saves you exactly the three hours where nothing else moves. The more we take AI for granted, the more the unglamorous question matters: what do I do when it’s just not there?

And yes — for a brief moment, it was almost peacefully quiet.


Sources: