Login failures and unloadable chats hit users worldwide for roughly two and a half hours on Saturday, in OpenAI's second major disruption this week.

OpenAI has restored ChatGPT after a global outage on Saturday, July 25, that left users unable to log in, load past conversations, or send new prompts across the web interface, mobile apps, the OpenAI API, and the Codex coding platform.

The disruption began at approximately 5:00 AM ET (09:00 UTC / 2:30 PM IST), with independent monitoring showing complaint volumes climbing from around 4:47 AM ET. OpenAI acknowledged the incident on its status page at 5:30 AM ET, saying it was investigating the issue for the affected services. The company later confirmed it had applied a mitigation and was monitoring recovery. Services were reported back to normal by around 7:29 AM ET, and OpenAI's status page showed all systems fully operational at the time of writing.

What users saw

The failure mode pointed to a backend problem rather than a front-end or CDN issue. Affected users reported:

• The ChatGPT sidebar was stuck on a loading animation, with previous conversations never rendering
• "Too many concurrent requests" errors when attempting to send a message
• Login and session failures, including users being unable to authenticate at all
• Chat history appears missing or empty after a successful sign-in
• Voice and audio-based interactions are failing to connect

Developers reported elevated latency and rejected requests against API endpoints, with as many as 12 API components flagged on OpenAI's status page during the incident. Codex users saw the same failures propagate into code review and automation workflows that depend on those endpoints.

Global in scope

Telemetry indicated the outage was not regional. Downdetector logged more than 3,000 reports during the peak, with clusters from the United States, Europe, India, Japan, and Australia. "Is ChatGPT down?" trended on Google search throughout the morning.

The disruption cut across subscription tiers. Free, Plus, and Enterprise users were all affected, as were business and education workspaces relying on ChatGPT for production work — a detail that undercuts the assumption that paid tiers get insulated from capacity-related failures.

Timeline

Second outage in three days

Saturday's incident follows a considerably longer disruption earlier in the week. An outage that began on July 23 and ran into July 24 produced elevated error rates across ChatGPT, Codex, and the OpenAI API for close to 24 hours before OpenAI declared it resolved. Separately, on July 21, the company confirmed a partial outage that took image generation offline in ChatGPT.

Three incidents inside five days are a notable cluster for a platform of this size, and it lands against a backdrop of steadily expanding product surface area — Codex, Sora, the Realtime API, and the recently introduced GPT-Live voice models all sit on shared infrastructure. OpenAI's own status page reports 99.67% uptime for ChatGPT components across April to July 2026, against 99.93% for the APIs and 99.98% for Codex, making the consumer-facing surface the least reliable part of the stack over that window.

No root cause disclosed

OpenAI has not published a root cause analysis for Saturday's outage, and the status page entry did not identify which subsystem failed. The combination of authentication failures and concurrency-limit errors is consistent with a problem in a shared session or request-routing layer rather than in model inference itself, but that remains inference rather than confirmation until OpenAI publishes a post-incident review.

For organisations that have built workflows on the OpenAI API, the week is a practical argument for fallback routing and graceful degradation. Neither the July 23 incident nor Saturday's was survivable through client-side retries alone, given that authentication itself was failing.

This story will be updated if OpenAI publishes further details on the cause.