74,000 reports. That’s how many complaints piled up on September 3, 2026, when ChatGPT and Codex both went dark for several hours in one of the biggest disruptions OpenAI has logged all year. OpenAI’s own status page flagged elevated errors. The coding tool went down with the chatbot, which means a lot of people watched their editor and their rubber duck die at the same moment.
And yet, as I write this on September 26, 2026, ChatGPT is fine. UptimeRobot’s probes reached chatgpt.com on September 24 and completed a full check without errors. Status trackers report no current problems. The last outage anyone logged before now was August 27, 2026, and it lasted about 49 minutes.
So which is the real story? Both. That’s the part worth sitting with.
The outage pattern is a rhythm, not an emergency
Look at what’s actually on record for 2026:
- February 4 — users reporting trouble accessing ChatGPT
- July 25 — a global outage that hit the web app, the mobile app, and the developer APIs, with Gulf News covering disruption in the UAE and worldwide
- August 27 — roughly 49 minutes of downtime
- September 3 — the big one, 74,000 reports, Codex included
That’s not a service falling apart. That’s a service at enormous scale having the kind of bad days that every service at enormous scale has. If you’ve run infrastructure, none of this shocks you. Four notable incidents across most of a year, with one genuinely painful multi-hour event, is unremarkable engineering reality.
What’s remarkable is the reaction. Every single time, the internet performs the same ritual: search “chatgpt down,” refresh Downdetector, post a screenshot, declare AI a house of cards. Then it comes back, and everyone forgets until the next one.
The July outage is the one people should have learned from
I keep coming back to July 25, because that incident took down the web app, the mobile app, and the developer APIs together. That combination is the tell. If your fallback plan for the web app going down is opening the mobile app, you don’t have a fallback plan. You have the same plan twice.
And if your product calls the API, your users experienced your outage, not OpenAI’s. They didn’t see a status page. They saw your app broken. You absorbed the reputational cost of someone else’s bad afternoon.
September 3 made the same point harder by taking Codex down alongside ChatGPT. Two tools, one blast radius. Anybody who thought they’d diversified by using both learned they hadn’t diversified at all.
What I actually tell people who build on this stuff
I review AI tools for a living, and the uncomfortable truth is that most teams building on model APIs treat provider uptime as somebody else’s problem right up until it becomes theirs. A few things that separate the teams who shrug off outages from the ones who spend the day writing apology emails:
- Run a second provider behind an abstraction layer. Not someday. Now, while nothing is broken. Swapping providers under pressure at hour two of an outage is how bad code ships.
- Degrade instead of dying. Cached responses, a queue that retries later, a plain message telling users what’s happening. Any of these beat a spinner.
- Know the difference between the app being down and the API being down. Monitor the endpoint you actually depend on, not the one you read about on social media.
- Don’t stack all your agent steps on one vendor. Agent chains multiply exposure. A five-step workflow on a single provider isn’t five chances to succeed, it’s five chances to fail.
The consumer side of this is simpler and bleaker
If you’re a subscriber rather than a builder, you have almost no recourse. You wait. You refresh. You maybe keep a second chatbot bookmarked so your workday doesn’t stop. That’s the whole toolkit. Paying monthly does not buy you a guarantee, and pretending otherwise just sets you up for a worse afternoon.
My honest read
ChatGPT’s reliability record in 2026 is boring in the way infrastructure should be boring, punctuated by a few genuinely rough days. Nobody should be writing think pieces about the death of OpenAI because of a 49-minute blip in August. But nobody should be building critical workflows on a single vendor either, and the September 3 numbers are a fair reminder of what that dependency feels like when it breaks.
The searches spike, the outage ends, the lesson evaporates. If you want to be in a different position next time, the work happens on the quiet days. Like today, apparently, since every monitor I can see says the thing is up and answering.
Use the calm. Build the fallback. The next incident is not a question of whether.
🕒 Published: