Three of the most widely used AI services stopped responding within the same window. Between late on 3 September and the early hours of 4 September 2026 (Japan time), OpenAI's ChatGPT, Anthropic's Claude and Grok each ran into trouble, and all three confirmed the problem on their own status pages. From the first sign of trouble to the final recovery notice, the episode lasted about four hours. Three vendors failing together is unusual — but the causes do not appear to reduce neatly to a single one.
Three services, minutes apart
Claude and Grok were the first to wobble, with connection problems reported from around 22:30 Japan time on 3 September. ChatGPT followed a little later, with symptoms surfacing around midnight.
For users this was less "slow" than "unreachable." Downdetector collected more than 35,000 reports, the vast majority of them tied to ChatGPT. Although it was the middle of the night in Japan, it was a weekday morning in the United States, so plenty of people had their work stop mid-task.
Anthropic marked the issue resolved and returned its status to normal at 01:16 Japan time on 4 September. OpenAI reported recovery at 01:55, and Grok at 02:07. All three converged within roughly one hour of each other.
The damage went well beyond the chat window
What stood out was how far the impact reached past the conversational interface.
OpenAI's status page listed 15 affected ChatGPT components plus four on the developer-facing Codex side. Conversations, login, search, file uploads, voice mode, image generation, Deep Research and agent features degraded more or less in unison. The company also noted that after recovery, some Codex users might need to pair their mobile device again.
Anthropic saw something similar. Elevated error rates were first reported on Mythos 5.1, Fable 5.1 and Opus 5, then widened to include Opus 4.8 and 4.6. The blast radius covered not only the Claude app but the API, Claude Code and Claude Cowork — meaning internal tools and automation built on Claude stopped at the same moment.
The chain extended further out. Cursor, the AI coding assistant, posted its own incident caused by the Grok and Claude outages. Any service that leans on a specific model provider goes down when its upstream does. Obvious enough in principle; on this day it happened with two providers at once.
When three vendors fail in the same window, it is tempting to look for common ground underneath. Yet the explanations published so far do not converge on a single root cause.
For OpenAI, the trigger is described as a routing fault that began around 07:43 Pacific time on 3 September. That sits inside the company's own traffic-control layer and is not something shared directly with the others. Anthropic and Grok, meanwhile, have not published a detailed technical breakdown.
During the incident, reports of trouble at a major cloud platform also spiked in a similar window, prompting speculation that shared infrastructure was involved. No vendor has confirmed that, and it remains inference drawn from circumstances rather than a stated cause.
The more interesting detail is that Google's Gemini never declared an outage. Downdetector reports for Gemini did rise in the same period, but that is consistent with users piling in after the other services failed, producing congestion and mistaken reports. That Google — which owns its data centres and its chips — stayed up is worth noting when thinking about how infrastructure is owned.
Designing for the assumption that it will stop
The episode shows that AI availability is no longer a single vendor's problem. If your team leans on agents such as Claude Code or Codex all day, work halts the moment they stop. Unlike a spreadsheet or an email client, there is no local fallback sitting on the machine.
Practical measures include keeping the ability to switch between models from different providers, queueing important jobs so they can be retried, and subscribing to each vendor's status page. Learning about an outage from a customer complaint is the situation worth avoiding.
That said, all three went down together this time. Even with a switchover target prepared, the chance that the target is also down in the same window is not zero. Redundancy is not a cure-all, and that is part of what makes this incident worth remembering.
Summary
Late on 3 September and into the early hours of 4 September 2026 Japan time, ChatGPT, Claude and Grok all suffered outages, recovering within roughly an hour of one another. The impact reached beyond chat interfaces into APIs, agent features and downstream services such as Cursor. OpenAI pointed to a routing fault, no shared cause has been published, and Gemini never declared an incident. The more central AI becomes to daily work, the larger the cost of downtime — and designing around failover and retries is quietly becoming the default.
