Three of the biggest names in artificial intelligence hit turbulence at the same time on September 3, 2026, when OpenAI, Anthropic, and Google all reported service disruptions affecting their flagship AI products.
The timing was notable. This was not one company having a bad day. All three stumbled in overlapping windows.
What actually broke
OpenAI bore the most documented impact. At approximately 10:58 AM UTC, the company detected elevated error rates across ChatGPT and Codex, affecting 15 components on ChatGPT and 4 on Codex. Engineers applied mitigations within about 24 minutes, bringing the incident to a close by 11:22 AM. A separate, shorter ChatGPT Work Mode disruption earlier that morning lasted roughly 5 minutes before resolving on its own.
Anthropic’s issues were technically the day prior. Claude Sonnet 5 experienced elevated errors on September 2, with that incident lasting approximately 25 minutes before engineers resolved it.
Google’s Gemini had it roughest in terms of duration. User-reported disruptions on September 3 coincided with at least one detected issue that took roughly 2 hours to fully resolve.
The affected services spanned consumer interfaces, developer APIs, image generation tools, and account creation flows.
Context: this is more common than the press releases suggest
Independent uptime monitoring services tracked a reassuring headline number: major AI providers maintained availability at or above 99.9% across 2026. That figure sounds airtight until you do the math. At 99.9% uptime, a service can still be down for nearly 9 hours over the course of a year.
Monitoring data also showed that AI providers collectively encountered dozens of similar outages in 2026 alone, most of them tied to scaling pressure on inference infrastructure.
Recovery times were swift across the board, measured in minutes to a couple of hours, and all three companies published status updates in line with modern transparency expectations.
What this means for businesses and the AI vendor landscape
The obvious response is multi-provider architecture, meaning companies design their AI-dependent workflows so that a failure at OpenAI can be absorbed by routing requests to Anthropic or Google, and vice versa. The catch is that AI APIs are not perfectly interchangeable. A prompt tuned for GPT-4o does not always produce equivalent output from Claude Sonnet 5 or Gemini. Building genuine redundancy requires engineering investment, not just a second API key.