What Happened During ChatGPT’s September 15 Disruption?

A surge of reports on Tuesday, September 15, prompted users to search Google for “Is ChatGPT down?” Downdetector had received more than 2,000 reports involving OpenAI by the reported afternoon peak, and one contemporaneous report said roughly 85% of those complaints concerned ChatGPT. That count represents reports submitted to an outage tracker—not the number or percentage of all ChatGPT users affected. 3537

What users saw—and what was confirmed

The available evidence supports a short-lived, user-reported access problem affecting at least some ChatGPT users. It does not establish that every ChatGPT user was affected, nor does it identify a root cause.

At the time reports were rising, OpenAI’s public status checker was reported as showing no known issue. 37 OpenAI’s official incident feed does show a separate minor incident affecting GPT-5.6 and GPT-5.6 Instant on paid plans from 07:12 to 07:53 UTC on September 15. However, the published information does not clearly connect that incident to the later report spike. 17

That distinction matters: a crowdsourced spike can reveal real friction before a provider declares an incident, while a provider’s status page may reflect only problems that have been verified, meet a reporting threshold, or affect a defined component.

Why status tools gave different answers

The September 15 signals were inconsistent, but not necessarily contradictory.

• Crowdsourced reports: Downdetector’s more-than-2,000 reports showed that many people were encountering problems at roughly the same time. About 85% were categorized as ChatGPT-related in one report. 35
• Official status: OpenAI’s public page was reported as clear during the surge, even though its incident feed separately records the earlier paid-plan model issue. 1737
• Third-party trackers: Entireweb labeled ChatGPT degraded on September 15, citing 140 reports over 24 hours and 11 in the preceding hour. 9 StatusGator listed the service as operational/resolved while also displaying 201 user-submitted reports over 24 hours. 16
• Synthetic monitoring: UptimeRobot said its North American probe could reach chatgpt.com without unusual response times or error codes. 3

These services are measuring different things. A homepage probe can succeed even when users are unable to log in, load a conversation, use a particular model, or receive responses. Conversely, user reports can surface localized or account-specific failures without proving a common infrastructure fault.

How it compared with the September 3 outage

September 3 was the clearer and much larger incident. Reports climbed above 5,000 by 7:53 a.m. PT, exceeded 22,000 about 10 minutes later, and were reported above 74,000 by 8:23 a.m. PT. 32 Other reporting put the ChatGPT count above 37,000 around 11 a.m. ET. 33

Unlike the September 15 spike, OpenAI acknowledged the September 3 problem. The company said a routing error beginning around 7:43 a.m. PT made ChatGPT and Codex unavailable for some users, and that a mitigation was implemented around 8:17 a.m. PT. 147

Claude and Grok also had overlapping disruptions that morning, while Gemini drew user-error reports. But a shared Microsoft Azure East US cause for all of those services was not confirmed by the vendors’ published explanations. Reporting noted no broad Azure outage confirmed to explain all of the failures. 4144

What does “58 incidents in 90 days” mean?

One independent tracker counted 58 ChatGPT incidents in the prior 90 days, including one major and 57 minor incidents. 27 That figure should be treated cautiously: third-party trackers can use different scopes and counting rules, and another service listing incidents sourced from OpenAI’s status page displayed 25 incidents. 21

The more useful conclusion is not a single headline count. OpenAI’s status history shows recurring incidents across distinct products and features—from ChatGPT and paid-plan models to Work Mode, image generation, file uploads, and API services. 1720

The practical reliability lesson

The September 15 report surge was enough to disrupt users and trigger widespread “is it down?” searches, but it was not enough to prove a universal outage or an Azure-related failure. OpenAI did not publish a technical explanation tying the reported afternoon spike to a specific cause.

For people and teams that rely on AI in daily workflows, the safest response is to monitor more than one signal: the vendor’s status page for confirmed incidents, an independent monitor for reachability, and crowdsourced reports for user impact. Keep an alternate tool or manual fallback for time-sensitive work—especially when a service may be reachable in one region or workflow while failing in another.