ChatGPT, Claude and Grok Went Down at Once. No Shared Explanation Yet.

A note on this update: during this correction pass, our reporting tools could not reach status.openai.com, status.anthropic.com, or the news outlets originally cited — every attempt returned a network access block. Everything below traces to the briefing packet assembled for this story, presented as sourced-but-not-independently-verified, not confirmed fact. We removed several specific claims we could not verify ourselves. See the corrections noted at the top of this story for exactly what changed.

Live updates

Newest first. Times below come from the briefing packet for this story. Its timezone was never stated and we have not been able to independently confirm one — treat every clock time here as approximate and unconfirmed-timezone, and check the live status pages in Sources below for the current picture.

• 2:58 PM (timezone unconfirmed) — OpenAI's status page shows an update on the incident: investigating.
• 2:56 PM (timezone unconfirmed) — OpenAI's status page shows the "Elevated errors across ChatGPT and Codex" incident window opening, marked Degraded performance / Investigating. 19 components are listed degraded across ChatGPT and Codex, including Conversations, Login, Search, Voice mode, Image Generation, Deep Research, and Codex in Desktop.
• Thursday, general reports — Reports spike for Claude and Grok alongside ChatGPT. Cursor's own outage is attributed to the Grok and Claude problems it depends on. Google's Gemini is reported as unaffected; Google hasn't confirmed an outage of its own.
• 12:04–12:10 AM (timezone unconfirmed) — A separate, shorter OpenAI incident, "ChatGPT Work Mode High Error Rates," opens and is resolved six minutes later.
• Wednesday, Sept 2 — An earlier OpenAI account-creation error incident is resolved.

ChatGPT, Claude and Grok all broke around the same time Thursday. All three companies indicated they were looking into it. As of this story's last edit, no company had published a shared, on-the-record explanation for why.

This wasn't three separate hiccups. It was three of the world's biggest AI chatbots, and the same signal from every company: investigating. Downdetector logged more than 20,000 US reports for OpenAI, according to the briefing packet for this story — we have not been able to independently confirm that figure, and outlets that track these outages often report different counts depending on when they check and how they filter reports, so treat any specific number here, including this one, as approximate. The trouble spread further than the chatbots themselves: Cursor, a coding tool that many developers use daily, had its own outage attributed to the Grok and Claude problems underneath it. Google's Gemini kept working, even though reports about it reportedly spiked too — Google never confirmed an outage of its own.

What caused it — nobody knows yet

No company has published a shared, on-the-record explanation for what caused this. Treat any specific answer you see online as a guess, including anything in this paragraph.

Here's what's worth watching, labeled as a lead and not a finding: Microsoft's Azure cloud reportedly saw its own spike in outage reports at the same time. OpenAI, Anthropic and xAI all run at least part of their operations on Azure, though other cloud providers, including Google, are part of the mix too. That timing lines up, per our sourcing material. It is not proof, we have not independently confirmed it, and no company has said Azure was the cause.

If it does turn out to be a cloud problem, it wouldn't be a strange one. Big cloud outages are rarely dramatic — historically, they're often one bad change, in one place, that happened to sit underneath a lot of the internet at once. We are not citing specific past incidents here because we could not verify the details ourselves this session.

The timing noise

Some of Thursday's online chatter connected the outage to rumors that OpenAI was about to announce a new model, code-named "Astra," possibly arriving as GPT-6 instead of the expected GPT-5.6. OpenAI and ChatGPT's own social accounts had posted teaser messages earlier the same day, and users joked the outage looked like an Apple Store going dark right before a launch.

That's speculation, not reporting. Nobody at OpenAI has connected the outage to any announcement, and we're not connecting them either. We're telling you the rumor exists because you'll see it today, not because we think it explains anything.

Why this isn't a Pentagon story

Four days before this outage, the Pentagon launched two new AI tools for military use on a platform called GenAI.mil: "ChatGPT Mil" from OpenAI and "Grok for Government," joining Google's Gemini there. Both are cleared to handle Controlled Unclassified Information at Impact Level 5, a serious security tier. More than 1.7 million of the Pentagon's roughly 3 million personnel have used the platform since it opened in December.

Some people jumped straight to a theory Thursday: the military rollout and the outage must be connected. It doesn't hold up. Military AI systems run in a separate, secured environment — a rollout inside that environment and an outage on consumer servers are two different systems entirely. There's no evidence tying the outage to the Pentagon rollout, and we're stating that plainly so the theory doesn't spread further from here.

The real story is simpler, and in its own way bigger: a huge amount of everyday AI use runs through three companies, and those companies lean on a small number of cloud providers. When one provider has a bad day, a lot of unrelated tools can go dark together. That's what happens when a young industry builds itself on a small number of pipes.

A note on status pages

Status pages need a human or an automated check to post an update, and both can lag behind what users are already seeing — a normal feature of a fast-moving incident, not a sign anyone is hiding something. An earlier draft of this story cited a specific example of a status page appearing to disagree with outside reporting; we removed that example because we could not verify it against a source we opened ourselves, and it may have come from a stale snapshot. We're not replacing it with an unverified example either. Check the live status pages in Sources below for the current picture, not what's written above.

What an "outage" actually is

Status page. A webpage each company runs to report its own system health, like status.openai.com. It's updated by hand or by automated check, so it can lag behind what users are actually seeing.

Degraded vs. down. "Degraded" means a service is slow, glitchy, or failing for some people. "Down" means it doesn't work at all. Thursday's incident was labeled degraded, though it felt like "down" for plenty of people trying to log in.

Why three companies break together. ChatGPT, Claude and Grok are separate companies with separate software, but none of them own the physical computers that software runs on — they rent that from cloud providers like Microsoft Azure, Google Cloud, or Amazon Web Services. If a shared piece of that rented infrastructure has a bad day, every company renting it can have a bad day at once, even though they're competitors with nothing else to do with each other.

What it means here in West Texas

By Thursday afternoon, a Lubbock roofing crew building an estimate in ChatGPT, an ag operation running planting numbers, a clinic rescheduling a full afternoon, and a dispatcher checking a route all hit the same wall: the tool was gone, no warning, no ETA.

That's not a reason to panic. Plan for it the way you already plan for a power flicker or a slow internet day.

• Know your fallback before you need it. If ChatGPT goes down mid-estimate, do you have a second model, or the manual process you used before AI? Decide that today, not while a customer is standing in front of you.
• Don't let one tool run your money. If AI drafts your invoices or answers customer texts, make sure a person can still do it by hand when the tool is out. One point of failure in billing or customer comms is a bad bet.
• Bookmark the real status pages. status.openai.com and status.anthropic.com beat a guess on social media. Check them before you assume it's your wifi.
• "Is it me or them?" Before you reboot your router or call your IT guy, check a status page or a public tracker. Fifteen seconds saves an hour of troubleshooting the wrong problem.

This is normal infrastructure behavior in a young industry, not a sign anything is broken about AI itself. Plan for it, and keep working.