AI Outage September 2026: ChatGPT, Claude and Grok

On September 3, 2026, OpenAI, Anthropic and xAI each ran an outage incident inside the same four hours. ChatGPT and Codex returned elevated errors across nineteen tracked components. Claude Opus 5 and Opus 4.8 kept failing after the rest of the Claude line had already recovered. Grok's models went down on the web and in both mobile apps. All three closed the same day, and no lab has published a shared root cause.

The coincidence got the headlines. The useful finding is underneath it: Cursor, which runs on all three, stayed broken for longer than any of the providers that broke it.

What the status pages actually recorded

Every figure below comes from a vendor status page, in UTC, as recorded on the day.

Anthropic opened on Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5, then widened the list at 13:50 UTC to include Opus 4.8 and Opus 4.6 (Claude status). OpenAI logged nineteen affected components, fifteen under ChatGPT and four under Codex, and posted three updates: investigating, mitigation applied, resolved (OpenAI incident record). xAI recorded a models outage on grok.com under incident INC25664c15.

Anthropic's timestamps line up with the reporting: its 13:41 UTC update is the 9:41 a.m. ET moment that outlets quoted all afternoon.

The agent tool was down longest, and it was down last

Cursor did not open one incident. It opened four (Cursor status):

1. 13:41 UTC, service degradation affecting all Grok models, Automations, Cloud Agents, Grok Bot and Review Agents. Resolved 17:07.
2. 14:17 UTC, elevated errors for Anthropic models. Resolved 16:31.
3. 15:17 UTC, elevated errors for OpenAI models. Resolved 17:05.
4. 17:42 UTC, degraded performance of Grok 4.6. Resolved 19:21.

Read the fourth one again. Cursor opened a new incident at 17:42 UTC, after every provider incident that day had already been marked resolved. Anthropic had closed at 16:23. OpenAI had closed at 16:55. The last upstream incident cleared at 17:07.

So the tool that sits on top of three labs carried 5 hours 40 minutes of degraded surface on a day when its worst-hit provider carried 3 hours 37 minutes. Failures do not divide across providers. They queue.

That is the number to take to an architecture review. Your agent's availability is not the best of its providers, and it is not the average. It is closer to the union of their bad days plus your own recovery time.

What did not come back when the API came back

OpenAI's resolution note carried one line that most coverage skipped: "Some Codex remote control users may need to pair their mobile device again following this incident."

The API recovered. The pairing did not. Somebody had to walk to a phone and redo a setup step before an agent that was already installed and already paid for would run again.

This is the gap between a service being up and your work resuming. A chat product recovers when the endpoint recovers, because its whole state is a text box. An agent holds state outside the model: a device pairing, an open session, a checked-out branch, a half-applied edit, a queue that drained into nothing while the loop was dead. None of that is on a status page, and none of it is restored when the incident closes.

Was it Azure?

Probably the most repeated claim of the day, and the least supported one.

Azure published no incident for September 3, 2026. Its status history page lists the most recent post-incident review as a July 23, 2026 West US connectivity event (Azure status history). Several outlets named Azure as the shared substrate because all three labs use it somewhere. No cloud provider confirmed it, and no lab named it.

Google published no matching Gemini incident, which is the one detail that makes the shared-infrastructure theory interesting rather than idle. It is still a theory.

Here is the honest sentence: three independent status pages went red inside four hours, the labs closed three separate incidents with three separate mitigations, and nobody has published a joint root cause. Waiting for one is not a plan. Designing as though the next one arrives without warning is.

Why this hit agents harder than chatbots

A chatbot outage costs you an answer. An agent outage costs you a run that was already in progress, and the cost scales with autonomy.

The autonomy level predicts what you lose:

A Copilot tool fails in front of a human who is already looking at it. A Fully Autonomous agent fails at a point nobody chose, in a repository or a shell, while the person who started it is in a meeting. Same outage, different recovery cost, and the difference is the autonomy value, not the quality of the tool.

One correction to a claim the category repeats: MCP-native does not mean provider-portable. Claude Code speaks the Model Context Protocol, so its tool servers survive a swap. Its model does not. The protocol makes your tools portable, not your inference, and those are two different layers. An MCP server that stayed up all afternoon is worth nothing if the only model wired to call it is the one returning errors.

Four decisions to make before the next one

1. Route to a different lab, not a different model from the same lab.

September 3 proved this rather than asserting it. At 15:25 UTC Anthropic reported that most Claude models had recovered and only Opus 4.8 and Opus 5 were still affected. A fallback chain of Opus 5 to Opus 4.8 would have failed twice inside the same incident. A chain that left the vendor would have cleared on the first hop.

2. Make the swap a configuration change.

If moving providers means rewriting a tool loop, you will not do it during an incident. This is the practical argument for a provider-agnostic framework over a single vendor's SDK, and it is worth the abstraction cost precisely on days like this one.

3. Decide per workflow what may pause and what must degrade.

Not every job needs a fallback. A nightly summarization job should pause. A customer-facing agent should degrade to a smaller or slower model and say so. Write the list before you need it, because during an incident every workflow looks urgent.

4. Write down what has to be re-paired, re-authenticated and re-run.

The Codex device-pairing note is the whole reason this step exists. Recovery of the provider is step one of your recovery, not the end of it.

What to look up after a day like this

Not "which chatbot is best". The question after September 3 is narrower: what does this tool do when its model is gone.

Three things worth checking on any agent before you depend on it. Whether it can be pointed at a second provider without a rewrite. What autonomy level it runs at, because that sets your blast radius. Whether it speaks a protocol or only a vendor's API, because the first survives a swap and the second does not.

Every listing in this directory records its autonomy, environment and ecosystem, plus whether it is MCP-compatible and what it costs. Start with the Claude Code listing if your agent lives in a repository, or the LangChain listing if you want the provider swap to be a line of config.

Vendors have bad days. Three of them had one together. The teams that felt September 3 as "chat was slow" will build the same way next quarter, and the teams that watched Codex, Claude Code and Cursor go quiet will start treating models as suppliers with an SLA they do not control.

Frequently asked questions

Why did ChatGPT, Claude and Grok all go down at the same time?

Nobody has said. OpenAI, Anthropic and xAI each opened separate incidents on September 3, 2026 and each closed its own without publishing a root cause. No joint post-incident review exists. Shared cloud capacity is a reasonable hypothesis and it is still a hypothesis.

Was the September 2026 AI outage caused by Azure?

Azure published no incident for September 3, 2026. Its status history page shows the most recent post-incident review covering a July 23, 2026 West US connectivity event. Several outlets named Azure as a likely shared cause, but no cloud provider and no AI lab confirmed it.

How long was the September 3, 2026 AI outage?

The provider incidents ran between 19 minutes and 3 hours 37 minutes. Anthropic's multi-model incident lasted 2 hours 57 minutes and OpenAI's lasted 2 hours 12 minutes. Cursor, which depends on all three, logged four incidents spanning 5 hours 40 minutes.

Was Gemini affected by the September 2026 AI outage?

Google published no matching incident. User reports for Gemini did spike on outage trackers during the same window, which is common when several large services fail together, but Google never confirmed a corresponding outage on its own status channels.

How do I stop a model outage from breaking my agent?

Route to a second lab, not a second model from the same lab. On September 3, Anthropic recovered most of the Claude line while Opus 5 and Opus 4.8 stayed broken, so an Opus-to-Opus fallback would have failed twice. Then write down what must be re-paired after recovery.