On Thursday 3 September 2026 outages hit OpenAI's ChatGPT and Codex, Anthropic's Claude and xAI's Grok within the same morning. The Register reported the sequence the same day. We rely on that report for the company statements, and we quote them as The Register quoted them.
What happened
All times are US Pacific unless noted.
- 6:30 am: xAI began investigating a Grok outage.
- 7:43 am: OpenAI says a routing error began affecting ChatGPT and Codex.
- 8:17 am: OpenAI said a fix was implemented and was being monitored.
- Anthropic reported a partial outage across Claude.ai, Claude Code, Claude Cowork and the Claude API. Service was restored at 16:16 UTC, after three hours and six minutes. The report does not give the start time.
- Later that day SpaceX apologized for the Grok outage and tied it to its Memphis compute center.
What is confirmed, reported and unknown
- Confirmed by the companies, as quoted: OpenAI named "a routing error." SpaceX named "an outage at our Memphis compute center." Anthropic said "an infrastructure issue caused a partial outage."
- Reported by The Register: all three companies use Cloudflare in different capacities. Cloudflare said it was not experiencing significant disruptions. AWS, Google Cloud and Microsoft Azure status pages showed no relevant problems, though The Register notes those pages do not always reflect issues promptly.
- Unknown: whether the three events share a cause. The article says SpaceX's statement "suggests one potential reason" for Anthropic's problems and does not confirm a link. Also unknown are the cause of OpenAI's routing error, the specific infrastructure issue behind Claude's outage, Grok's total duration, and user counts.
The honest reading is that three outages overlapped. Anyone who writes that one provider caused all of them is going beyond the record.
What to do now
- List every feature in your product that calls a hosted model API, and name the provider for each.
- For each, decide the fallback: a second provider, a cached answer, a queued job or a plain error message.
- Set client timeouts and retry budgets so a provider outage does not exhaust your own threads.
- Check your incident log for 3 September and measure the real customer impact.
- Do not assume two providers are independent. Ask where each hosts its compute and what edge network it uses.
How this could have been avoided
You cannot prevent a provider outage. You can limit what it does to you.
- A second model path. Outages here were staggered and separate. A tested failover to another provider would have covered most of the window.
- Graceful degradation. Decide in advance which features switch off and which fall back to non AI behavior.
- Circuit breakers and timeouts. A hung model call should not hold a request for minutes.
- Dependency mapping. Shared infrastructure behind two vendors is a single point of failure you only see if you ask.
Try it yourself
Draw your AI dependencies on one page: feature, provider, fallback. Mark any feature with no fallback in red. That list is your first resilience backlog.
Click through it at your own pace, or open it full screen.
Quick check
Was Cloudflare the common cause?
Not confirmed. All three companies use Cloudflare in different capacities, but Cloudflare said it had no significant disruption, and no company named it.
What did OpenAI say caused its problem?
A routing error starting around 7:43 am Pacific. The report does not say more.
