1300 166 138 Tell us what you need
News / AI Agents

Claude and ChatGPT went down together for 100 minutes

By Monika Tantau • Published 2 months ago
Claude and ChatGPT went down together for 100 minutes

Key takeaways

• Anthropic logged major impact on Claude Mythos 5.1, Fable 5.1, Opus 5, Opus 4.8 and Opus 4.6
• OpenAI logged elevated errors across 15 ChatGPT components and four Codex components
• Both providers had an open incident for 100 minutes, from 14:43 to 16:23 UTC
• In Melbourne that window ran from 12:43am to 2:23am, so overnight automations wore it

Most Read

Anthropic and OpenAI both ran degraded services on 3 September, and for 100 minutes the two largest frontier providers had an open major incident at the same time. Anthropic reported elevated errors across every model in its top tier. OpenAI reported elevated errors across ChatGPT and Codex.

Neither company called it a full blackout, and neither one was. Both were error rate events rather than a dark screen. For anything that calls these models without a human watching, the practical result is the same, and in Australia that window fell in the middle of the night.

What the status pages record

Anthropic opened an incident at 13:26 UTC for Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5, then widened the list twenty minutes later to include Mythos 5, Fable 5, Opus 4.8 and Opus 4.6. By 15:25 only Opus 5 and Opus 4.8 were still affected. The company recorded impact as ending at 16:16 UTC and closed the incident at 16:23, on its own status page. A separate Claude Sonnet 5 incident had already come and gone between 12:37 and 12:56.

OpenAI opened its incident at 14:43 UTC, listing 15 affected ChatGPT components and four Codex components, applied a mitigation at 15:17 and marked the issue resolved at 16:55, according to its status page. It also warned that some Codex remote control users would need to pair their mobile device again afterwards.

Line the two timelines up and the overlap is exact. From 14:43 to 16:23 UTC, both vendors had an unresolved major incident on the board at the same moment. That is the 100 minutes.

It was not a shared supplier

It was not a shared supplier

The obvious theory is a common dependency underneath both, and the public record does not support it. Cloudflare logged nothing in that window at all. Its incidents that day were an increase in errors in Hong Kong between 09:21 and 09:36 UTC, and 522 errors in Seattle between 06:30 and 07:38, both closed hours before either model provider reported anything.

Amazon Web Services published no global event for the period either. On the evidence available, two separate companies had two separate bad afternoons that happened to land on top of each other. That is a less satisfying answer than a single point of failure, and a more uncomfortable one, because it cannot be designed around.

Why Australian teams barely noticed

Converted to Melbourne time the incident ran from 12:43am to 2:23am on 4 September. Almost nobody in an Australian office was at a keyboard, which is why the story arrived here as news rather than as a support queue.

The exposure sits somewhere else. Overnight is when Australian businesses run the work they deliberately moved out of office hours: batch document processing, scheduled reconciliations, enrichment jobs, the agent that drafts the morning summary before anyone opens a laptop. Those jobs fail quietly at 1am and are discovered at 9am, if they are discovered at all.

The same applies to anything customer facing that stays up while the team sleeps. A support assistant that answers at 2am is only useful if it fails in a way somebody sees, which is the least glamorous and most valuable part of any AI chatbot build.

What a fallback actually needs

The standard advice after a model outage is to add a second provider and route around the first one. This event is a useful test of that advice, because for 100 minutes the second provider was also degraded. A failover from Claude to ChatGPT during that window would have moved the problem, not solved it.

What survives an event like this is duller: a retry policy with proper backoff, a queue that holds work instead of dropping it, an alert that fires on failure rate rather than on a dashboard nobody is watching at 1am, and a fallback tier that includes a smaller model or a deferred run. Anthropic’s own timeline shows why tiering helps, since Sonnet 5 recovered while the larger models were still failing.

It is also worth knowing which model your workload actually pins to, and whether it can accept a downgrade. If you have never made that comparison, our breakdown of Claude and ChatGPT for business workloads covers where each one is genuinely substitutable and where it is not.

Two vendors, one window

Frontier model incidents are not rare. Anthropic alone logged four in the three days to 3 September, most of them minor and short. What made this one worth recording is the overlap, and the fact that the usual mitigation was unavailable for the duration.

The last time Claude went dark for Australian users it was a policy decision rather than a technical one, and we tracked the export order that pulled Fable 5 at the time. Different cause, same lesson: the resilience question is not which vendor you pick, it is what your systems do in the hour when the answer does not come back.

Scroll to Top