What happened
On Thursday morning, a rare, overlapping AI outage hit four major providers at once — OpenAI's ChatGPT and Codex, Anthropic's Claude, xAI's Grok, and Google's Gemini all degraded within the same few-hour window. Ars Technica called having all four go down together "practically unheard of."
Anthropic reported trouble first. At 9:23 am Eastern, the company logged a "partial outage" with elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5. Anthropic said it had identified the cause about 15 minutes later, deployed a fix, and marked the incident resolved by 12:16 pm. A separate, second incident then hit Claude Sonnet 5 with elevated errors just after noon.
OpenAI followed at 10:43 am, reporting "elevated errors across ChatGPT and Codex" causing degraded performance. A mitigation went live roughly 30 minutes later, and the issue was marked resolved by 12:55 pm.
xAI's Grok displayed a user-facing message saying the model "is experiencing issues" while the company worked on a fix. DownDetector reports for Grok jumped from fewer than 10 just before 9 am to 1,365 by 9:45 am, before easing back to 273.
Google never publicly acknowledged a Gemini problem, but DownDetector reports for Gemini spiked from 23 around 10:30 am to 412 shortly after 11 am, and monitoring service StatusGator flagged a "likely outage" for the Gemini API between 10:45 and 11:15 am.
### Cloud infrastructure stayed mostly clear
Amazon Web Services, Microsoft Azure, and Cloudflare reported no major incidents that morning, though DownDetector reports ticked up somewhat for all three — a sign the disruption sat in the model layer, not the underlying cloud infrastructure.
Why it matters
Isolated AI outages are routine and usually go unnoticed outside the affected user base. Anthropic reports 99.4% uptime for Claude over the past 90 days, with its last comparable incident a three-hour partial outage on August 24. OpenAI reports 99.63% uptime for ChatGPT and 100% uptime for Codex over the same period, with an earlier stretch of elevated latency in "Work Mode" on August 31.
What made September 3 different is that four independent providers, running on separate infrastructure, degraded inside the same 90-minute window. One theory, floated on X by OpenAI's Thibault "Tibo" Sottiaux, is a "domino effect": when one provider slows down, users and automated systems fail over to backup models, and the sudden traffic surge stresses those backup providers in turn.
For any business that has quietly built a single point of failure — one chatbot, one API, one coding assistant — into daily operations, this is the wake-up call. A single-vendor AI stack is now a business-continuity risk in the same category as a single-region cloud deployment, and it deserves the same level of planning.
How to use it today
Multi-provider redundancy needs to be a running default, not a plan drafted after the next outage. In practice:
- Keep at least one backup model provider configured and tested ahead of time, not just bookmarked for later.
- For customer-facing workflows — support bots, content generation, code review — build a fallback path that swaps providers automatically once error rates spike, instead of letting the whole workflow fail.
- Track each provider's own status page alongside a third-party monitor like DownDetector or StatusGator, since companies don't always self-report; Google stayed silent on Gemini even while its own dashboard showed a spike.
For teams that don't want to manage multiple paid API keys just for redundancy, free browser-based tools are a practical stopgap. A site like [mykreatool.com](https://mykreatool.com) bundles several no-cost AI tools, so a single provider outage doesn't stop a marketer or creator from finishing a task mid-deadline.
Who benefits
Marketers, support teams, and solo creators running lean AI stacks are the ones most exposed — and most helped by planning ahead now. Enterprises with dedicated infrastructure teams already build multi-model routing into their systems; small teams and freelancers usually don't, and an hour of downtime during a client deadline is where the real cost lands. Developers using Codex or Claude for active coding work are also directly exposed, since an outage there stalls in-progress work, not just a chat window.
Risks
### Overcorrecting into complexity
Wiring three or four providers into every workflow adds cost, latency, and maintenance overhead most teams don't need for occasional chat use. Reserve automatic failover for production-critical paths, not every internal task.
### Trusting a single status page
Google's silence on Gemini during a documented DownDetector spike shows provider status pages aren't always reliable in real time. Cross-check with an independent monitor before assuming a service is fine — or down.
### The "domino effect" is still unconfirmed
Sottiaux's explanation is a plausible theory from someone at OpenAI, not a confirmed root cause published by any of the four companies. Treat it as a hypothesis until providers release full postmortems.
Conclusion
Four major AI providers — OpenAI, Anthropic, xAI, and Google — went down within the same few hours on September 3, 2026, an overlap Ars Technica calls practically unheard of. Individually, each outage was ordinary: Anthropic resolved its incident in under three hours, OpenAI in about two. Together, they exposed how much daily work now depends on a handful of cloud AI providers with no built-in redundancy. The fix isn't panic — it's a tested backup plan, a couple of independent monitors, and, when budget is tight, free tools to bridge the gap until the primary provider comes back online.
Comments 0