It’s 9:14 on a Tuesday morning. You’ve got a half-written client email, three bullet points that need turning into paragraphs, and a meeting in twenty minutes. You open the tab you always open, type your prompt, hit enter, and get a spinning circle that never resolves. You refresh. Same thing. You check your wifi, which is fine. And then comes that small, strange jolt of panic: wait, how did I used to do this?
If you’ve had that moment, you’re in very large company. And if you’re having it right now, the first useful thing I can tell you is this: as of late September 2026, ChatGPT is running normally. Monitoring services checking chatgpt.com report clean results, with an automated check on September 24, 2026 at 19:53 GMT turning up nothing unusual. So if it’s broken for you today, the problem is more likely local than global.
The outages that actually happened
ChatGPT going down isn’t imaginary, though, and it isn’t rare. Looking at just this year:
- February 4, 2026 — widespread access problems significant enough to generate its own wave of coverage and commentary.
- July 25, 2026 — a global outage reported by Gulf News that hit users in the UAE and elsewhere. This one took out the web app, the mobile app, and the developer APIs at the same time.
- August 27, 2026 — a shorter disruption, roughly 49 minutes.
- September 3, 2026 — the big one. ChatGPT and Codex, OpenAI’s coding tool, both went down for several hours. Reports topped 74,000. OpenAI’s own status page flagged elevated errors. It stands as one of the largest disruptions the company has logged this year.
Notice the pattern in that July event especially. Web app, mobile app, and APIs all failing together tells you something about how these systems are built. You’re not using three separate products that happen to share a logo. You’re using three doors into the same building. When the building has a problem, every door is stuck.
Why this matters more than it used to
A couple of years ago, ChatGPT being unavailable was an inconvenience. You’d shrug and write the email yourself. What’s changed is how much people have built on top of it, often without quite realizing they were building.
The September 3 outage taking down Codex alongside ChatGPT is the clearest illustration. Developers weren’t just losing a chat window; they were losing part of their working setup mid-task. And when APIs go down, the effect ripples outward to every app, tool, and internal workflow that quietly routes requests through OpenAI in the background. Some of those users never knew they were ChatGPT users at all. They just knew their support tool stopped summarizing tickets.
This is the part I find most interesting as someone who explains AI agents to people who don’t build them. An agent is only as available as the model behind it. If you’ve set up something that reads your inbox, drafts replies, and files things away, that thing has a single point of failure it inherited from a company you don’t work for.
What to do when it goes dark
Practical steps, in order:
- Confirm it’s not you. Check OpenAI’s status page or a third-party monitor. Two minutes of checking beats twenty minutes of restarting your router.
- Don’t refresh in a panic. If it’s a real outage, hammering the page does nothing. The August event lasted under an hour. Some resolve faster than the time you’d spend troubleshooting.
- Keep a backup route. Another assistant, a different tool, or simply a rough draft you write yourself. Not because the alternative is better, but because having one turns a crisis into an annoyance.
- Know which of your tools depend on it. If you’ve wired up automations, write down what breaks when the model is unreachable. Doing that audit on a calm day is much nicer than discovering it live.
- Save your work as you go. Long conversations you care about deserve a copy somewhere that isn’t a browser tab.
The honest takeaway
Four notable disruptions in a year, one of them hours long with 74,000 reports, is not a scandal. It’s what running enormous infrastructure at enormous scale looks like. Electricity grids fail. Airlines cancel flights. Cloud providers have bad afternoons.
What’s new is emotional rather than technical. The spinning circle exposes how quickly a tool becomes a habit, and how a habit becomes a dependency without asking permission. The fix isn’t to use these tools less. It’s to notice where you’ve handed over something you’d want back on short notice, and to keep a plan B that costs you nothing to maintain.
Right now, the service is up. Good moment to figure out what you’d do if it weren’t.
🕒 Published: