OpenAI reports degraded performance in September incident affecting millions

OpenAI’s systems are back to full operational status following a service degradation on September 3 that knocked out key functionality across ChatGPT and Codex. The disruption affected at least 15 components, ranging from basic conversations and logins to search, file uploads, and image generation.

The timing was particularly inconvenient. OpenAI had been actively teasing the launch of its GPT-6 Astra model, building anticipation among its user base right as the infrastructure underneath started buckling.

What happened on September 3

The main incident kicked off around 2:56 PM UTC, when OpenAI’s status page flipped to “degraded performance.” Error rates spiked across the platform, locking users out of core features that millions rely on daily for work and development tasks.

It wasn’t even the first hiccup of the day. Earlier that same morning, a separate incident caused high error rates within ChatGPT’s Work Mode. That one was comparatively minor, resolved in roughly five to six minutes by 12:10 AM UTC.

OpenAI launched an investigation promptly and ultimately restored all systems to working order. The company confirmed through its communications that everything was back online, though the duration of the full restoration window wasn’t specified in its public updates.

A pattern that’s getting harder to ignore

OpenAI has logged hundreds of incidents since early 2025, a track record that’s starting to look less like growing pains and more like a systemic challenge.

The GPT-6 Astra wrinkle

The outage landed on the same day OpenAI was building buzz around its upcoming GPT-6 Astra model. There’s no public indication that the teasing of GPT-6 Astra was directly related to the infrastructure issues. OpenAI has refrained from explicitly correlating the outage with the launch preparations, keeping communications focused on operational transparency and resolution efforts.

OpenAI’s post-incident management appears to have been effective in this case, with systems returning to full operational status.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.