ChatGPT, Claude and Grok outages expose a bigger reliability problem for AI users
Three major AI platforms suffered service disruptions within hours of one another on Wednesday, September 3, turning a normal workday into a reminder that even the biggest cloud-based AI tools can suddenly become unavailable.
OpenAI reported an incident affecting ChatGPT and Codex at 14:58 UTC, describing it as elevated errors across multiple services. According to the company’s own incident log, 15 ChatGPT components and four Codex components were impacted. The issue was marked resolved at 16:55 UTC.
Anthropic also recorded two separate Claude incidents on the same day. The first affected Claude Sonnet 5 between 12:37 and 12:56 UTC. The second lasted much longer, running from 13:26 to 16:16 UTC, and involved several models, including Mythos and Fable 5.1, Mythos and Fable 5, Opus 5, Opus 4.8 and Opus 4.6.
xAI logged a Grok outage as well. Among the four major AI assistants, Google Gemini was the only one without an official outage notice that day, though user reports still appeared. Downdetector showed around 500 reports for Gemini, while ChatGPT peaked at more than 35,000 reports in the United States. Claude reached roughly 1,400 reports, and Grok climbed to around 1,200.
The outages themselves are now over. But the bigger issue is not a single bad day. The more important question is how reliable these AI services really are when measured over time.
The answer is available in the providers’ own reliability figures, although most users rarely check them.
OpenAI lists ChatGPT availability at 99.64 percent over a rolling 90-day period. Its APIs perform better at 99.94 percent, while Codex is listed at 100 percent. Anthropic reports 99.4 percent availability for Claude’s consumer service over the same type of 90-day window. Its API sits at 99.5 percent, Claude Code at 99.44 percent and Claude Cowork at 99.43 percent.
Those percentages may look high at first glance, but they become more meaningful when converted into time. Over 90 days, 99.64 percent availability equals roughly 467 minutes of downtime, or just under eight hours. Anthropic’s 99.4 percent for Claude works out to around 778 minutes, which is a little over 13 hours.
In other words, a service can look highly reliable on paper while still being unavailable for the equivalent of a full working day, or more, across a single quarter.
OpenAI’s incident history also shows that problems are not rare. Its public logs listed 36 individual incidents in July, 19 in August and five more during just the first three days of September. That makes the September 3 outage less of a strange exception and more of a visible example of a regular operational challenge.
What made this particular day stand out was the timing. ChatGPT, Claude and Grok all experienced trouble in the same general window, making the outage feel larger and more coordinated to users who rely on multiple AI tools.
Some observers pointed to Microsoft Azure’s East US region as a possible shared cause, since ChatGPT, Claude and Grok all depend on large-scale cloud infrastructure. However, that explanation has not been officially confirmed. Microsoft’s own published incident history did not list an East US event for that day, and neither OpenAI nor Anthropic named Azure as the cause in their incident notes.
That distinction matters. Cloud infrastructure is complex, and several services can fail around the same time for reasons that are not publicly visible. Without a confirmed post-incident explanation, it is not accurate to treat the Azure theory as settled fact.
Anthropic’s silence on the cause is especially notable because the company had been more direct only days earlier. On August 28, it stated that an issue with an upstream cloud provider had affected Claude Cowork and Claude Code on the web. For the September 3 incident, no similar attribution appeared.
OpenAI did add one practical detail for Codex users. Those who use Codex through mobile remote control may need to pair their device again after the incident.
For everyday users, the lesson is simple: when ChatGPT, Claude, Grok or Gemini suddenly stops working, the problem may not be your browser, your account or your internet connection.
Before clearing cookies, switching devices or reinstalling apps, it is worth checking the official status pages for OpenAI, Anthropic, xAI and Google. These dashboards show whether a provider has acknowledged an active issue and often include incident timelines once the problem is resolved.
However, official status pages have a limitation. They are maintained by the companies themselves and often lag behind real-time user reports. On September 3, outage reports surged on Downdetector before some official incident notices appeared. For a clearer picture, users should compare both sources: the provider’s status page and independent user-reporting platforms.
The outage also raises a bigger question for people who depend on AI tools for work: is subscribing to more than one AI service enough protection?
Not always.
Having both ChatGPT and Claude, or adding Grok as a backup, may seem like a safe strategy. But September 3 showed that multiple AI services can degrade at the same time. If providers share cloud infrastructure, regional capacity or other upstream dependencies, users may not know until everything starts failing at once.
A second subscription only helps if the second service is independent enough from the first. From the outside, that is difficult to verify.
For users who need guaranteed access to AI assistance, local AI models are becoming a more practical backup. These models run directly on personal hardware and do not require an active connection to a cloud AI provider. They are not yet equal to the largest online models in reasoning, coding or writing quality, but they can still be extremely useful when a cloud-based chatbot is unreachable.
Smaller language models can now run on consumer devices, including powerful phones and PCs. Meta’s Muse Glimmer is one example of a free model option that can run with 24 GB of graphics memory. While such tools will not fully replace ChatGPT, Claude, Grok or Gemini for every task, they can provide a dependable fallback during outages.
The September 3 AI outages were temporary, but they highlighted a lasting issue. AI assistants are becoming part of daily work for developers, writers, researchers, students and businesses. As reliance grows, downtime becomes more than an inconvenience. It becomes a productivity risk.
Cloud AI services will continue to improve, but no platform can promise perfect availability. Users who depend on these tools should know where to check service status, understand that official updates may be delayed and consider a local AI backup for critical moments.
The main takeaway is clear: ChatGPT, Claude, Grok and Gemini are powerful, but they are not immune to downtime. If AI is now part of your workflow, outage planning should be part of it too.
