Claude Goes Dark: Anthropic's Chatbot Hits the Skids Across Its Newest Models

Key points

• Anthropic's Claude assistant started throwing errors at around 9:41 AM ET on 3 September 2026.
• The outage hits six models: Mythos 5.1, Fable 5.1, Mythos 5, Fable 5, Opus 5, Opus 4.8 and Opus 4.6.
• Anthropic said at 10:49 AM ET it had identified the cause but was still working on a fix.
• Users are seeing failed requests and error responses when they try to chat with the affected models.
• No security incident has been claimed or confirmed. This looks, so far, like a plain reliability failure.

Claude is having a bad morning.

Anthropic, the company behind the Claude chatbot, confirmed on its official status page that users are hitting a wall of errors when they try to send requests to several of its flagship models. The trouble started at roughly 9:41 AM Eastern Time on 3 September 2026 and, at the time of writing, is still going.

If you use Claude for work, this is the outage where nothing loads and the little error message keeps coming back. If you don't, it's a useful reminder that a huge slice of the software people now depend on runs through a handful of AI providers, and when one of them sneezes, a lot of dashboards go red.

Which models are broken?

Six Claude models are affected, spanning Anthropic's newest releases and a couple of older ones people still lean on in production.

Anthropic first flagged elevated errors on three models, then a few minutes later expanded the list. The full set, per the company's own updates, is below.

Requests to these models are either failing outright or returning errors instead of a normal response. Applications built on top of Claude, think customer support bots, coding assistants, internal search tools, will inherit that same failure.

What actually caused it?

We don't know yet. Anthropic has said it identified the underlying issue but hasn't shared what it is.

By 10:49 AM ET the company confirmed engineers were still working on a fix. That's the entire public detail. No mention of a security incident, no mention of an attack, no mention of a bad deploy. It could be an internal configuration change, a capacity problem, or an infrastructure fault at a cloud provider. Until Anthropic publishes a post-incident writeup, anything more specific is guessing.

This was first reported by BleepingComputer as a straight availability story, and that's still the honest framing. A model going down is not, by itself, a security event. It becomes one only if the cause turns out to be malicious, or if the fallback behaviour of the apps built on Claude does something unsafe (say, silently switching to a less-restricted model with different safety filters).

Should users and businesses be worried?

Worried, no. Inconvenienced, yes, and it's worth thinking about what your tools do when the AI behind them stops answering.

For ordinary users, the fix is to wait. Retrying may or may not help while Anthropic is still working the issue. For businesses that have wired Claude into a product, this outage is the useful test case: does your app fail loudly and clearly, or does it hang, retry forever, or quietly hand the request to a different model without telling anyone?

A few practical things to check while Claude is down:

1. Are your Claude API calls timing out cleanly, or piling up?
2. If you have a fallback model configured, is it one you actually trust for the same task?
3. Are error messages reaching the humans who need to see them, not just the logs?

We'll update this piece when Anthropic publishes a cause. For now: not a breach, not a poisoning, not a jailbreak. Just a very visible reminder that AI infrastructure is still infrastructure, and infrastructure breaks.