ChatGPT and Codex outage resolved after a routing error, OpenAI says

A routing error, and a fix in about half an hour

OpenAI resolved an incident that made ChatGPT and Codex unavailable for some users on Thursday, September 3. According to the company's account, reported by WIRED and The Register, a routing error began around 7:43 a.m. PT, and a solution was implemented at about 8:17 a.m. PT. OpenAI spokesperson Kathleen Chaykowski gave both outlets the same wording: "A routing error starting around 7:43 am PT on Thursday, September 3 made ChatGPT and Codex unavailable for some users across platforms. As of about 8:17 am PT on Thursday, a solution was successfully implemented and is continuing to be monitored."

What users saw

The company's incident page listed elevated errors across ChatGPT and Codex, spanning fifteen ChatGPT components and four Codex components — from sign-ins and conversations to file uploads, voice mode, search, deep research and image generation. Downdetector, the third-party outage tracker, recorded a sharp spike in user reports during the window. OpenAI's final update marked the incident resolved at 16:55 UTC and noted that some Codex remote-control users might need to pair their mobile devices again.

Same day at Anthropic and xAI, with no confirmed shared cause

The disruption overlapped with incidents at two other AI providers, which fuelled speculation about a common dependency. Anthropic's Claude had a partial outage affecting several models and services, with impact declared over at 16:16 UTC. xAI's Grok was reported unavailable on the web, Android and iOS before recovering later in the day; SpaceX, xAI's parent company, linked the Grok disruption to an outage at its Memphis compute centre. Reporting has found no confirmation that the three incidents shared a cause: Cloudflare said it had no significant service disruption, and the status pages of AWS, Google Cloud and Microsoft Azure showed no relevant problems at the time.

OpenAI has not published a detailed postmortem of the routing failure. The outage came shortly before the expected launch of the company's next major model, Astra, though there is no evidence connecting the two. The practical record for anyone who depends on these tools is the one the status pages left: a fault that lasted roughly half an hour before a fix, with the incident formally closed at 16:55 UTC.