Anthropic's own status page confirmed on Thursday afternoon that Claude was having a rough stretch: elevated errors started hitting claude.ai, the Claude API, Claude Code, and Claude Cowork at 13:26 UTC, and by the time this went to press the company had found a cause but not yet shipped a fix. The incident spread fast on Hacker News, racking up 55 points and 34 comments within its first half hour, well before Anthropic had said anything beyond "we are investigating."

Here's what status.claude.com actually says happened, in the company's own words and timestamps, not a paraphrase pulled from social media chatter.

RelatedAnthropic Scales Project Glasswing to 150 Orgs in 15 Countries

  1. 13:26 UTCAnthropic opens the incident: "investigating elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5."
  2. 13:41 UTCStatus updated to "identified": the company says it has found the cause and is "working on a fix," without publishing what that cause is.
  3. 13:50 UTCThe affected-model list grows to five: Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6, described on the page as "an exhaustive list."

Four products are named as impacted: claude.ai itself, the developer-facing API at api.anthropic.com, the Claude Code CLI, and Claude Cowork. No error-rate figures are published anywhere on the page, which is thinner than the kind of post-incident detail AWS or Cloudflare typically post once a fix ships.

Which Claude products and models are hit by the Sept 3 incident Diagram showing four affected products, claude.ai, the Claude API, Claude Code, and Claude Cowork, feeding into five affected model versions: Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6, all shown in orange. A separate grey box shows Sonnet and Haiku model families, not named on Anthropic's exhaustive list and therefore not confirmed affected. STATUS.CLAUDE.COM — SEPT 3 INCIDENT claude.ai Claude API api.anthropic.com Claude Code Claude Cowork ELEVATED ERRORS, ROOT CAUSE IDENTIFIED, FIX IN PROGRESS CONFIRMED AFFECTED MODELS Opus 5 Opus 4.8 Opus 4.6 Fable 5 Fable 5.1 NOT ON THE "EXHAUSTIVE LIST" Sonnet family Haiku family genztech.blog
Fig 1 Anthropic's status page names five model versions across four products as affected. Sonnet and Haiku aren't on the "exhaustive list," which is the closest thing to an official all-clear those model families have right now.

What does "elevated errors" actually mean here?

Anthropic hasn't said. The status page walks through three updates in under half an hour, investigating, identified, exhaustive list, but at no point does it publish an error rate, a percentage of failed requests, or a technical explanation of what broke. That's a thinner public post-mortem than what AWS, Cloudflare, or GitHub typically publish once an incident like this resolves, and it leaves developers guessing whether they're seeing occasional 500s or a near-total outage on the named models.

Why does this reach past claude.ai's chat window?

Because two of the four named products, Claude Code and Claude Cowork, are the tools developers wire directly into their own build pipelines and agent workflows. A chat outage is an inconvenience. An API and CLI outage in the middle of a workday means CI jobs calling Opus 5 or Fable 5.1 start failing silently, and teams that built retry logic assuming occasional network blips, not a company-wide model outage, are the ones who feel this the hardest. GenZTech's own coding-model leaderboard tracks several of the exact model versions named in this incident, which is a reminder that "the model" and "the API that serves it" are two different failure points, and today it's the second one.

Which Claude models are still working normally?

Sonnet and Haiku. Anthropic's own wording calls its list of five affected models "exhaustive," which is about as close to an official all-clear as those two families get without a dedicated statement. If a workflow can tolerate a temporary model swap, routing traffic off Opus or Fable and onto a Sonnet-class model is the most concrete mitigation available right now, not a guess.

RelatedAnthropic watermarks every Claude output, worldwide

Who gets hit hardest by an outage like this?

Anyone running production software on top of api.anthropic.com without a fallback provider. Anthropic doesn't publish uptime SLAs or historical incident frequency on the same page, so there's no public baseline to say whether today's outage is typical or unusual for the company, but the concentration risk is the same regardless: teams that built exclusively on one foundation-model vendor, with no secondary provider to fail over to, have no lever to pull except waiting.

What to watch
  • The fix. Status was still "identified," not "resolved," as of this writing. Anthropic hasn't said how long a fix will take.
  • A post-incident writeup. Whether Anthropic follows up with a root-cause explanation, the way major cloud providers do after a widely-noticed outage, or leaves the incident log as the only record.
  • Sonnet and Haiku staying off the list. If either family gets added later, that's a sign the problem is broader than the current five-model scope suggests.

Our take

The bigger story here isn't the outage itself, providers have bad hours, that's not news on its own. It's how little public information ships with it. A five-model, four-product incident with zero error-rate data and zero technical explanation is a status page built to satisfy a support inbox, not developers trying to decide whether to page their on-call. As more teams route production traffic and agentic coding tools through a single model provider, the gap between "we're on it" and an actual post-mortem is going to matter more, not less.

Primary sources

Original analysis by GenZTech Team, based on Anthropic's live incident report. Source: status.claude.com.