September 2026 AI Tools Outage & Uptime Report
We continuously monitor 18 major AI tools including ChatGPT, Claude and Gemini. A recap of incidents recorded on providers’ official status pages in September 2026.
Longest outage: Anthropic（99h 21m）
By provider
| Provider | Incidents | Downtime | Worst state | Uptime |
|---|---|---|---|---|
| Cursor | 10 | 26h 54m | Major outage | 96.26% |
| Anthropic | 8 | 108h 30m | Partial outage | 84.93% |
| OpenAI | 5 | 8h 27m | Major outage | 98.83% |
| GitHub | 4 | 14h 37m | Major outage | 97.97% |
| ElevenLabs | 3 | 3h 50m | Partial outage | 99.47% |
| Perplexity | 2 | 2h 59m | Major outage | 99.59% |
| Stability AI | 1 | 4h 8m | Partial outage | 99.43% |
Incident counts and downtime cover only "major-or-worse and resolved" incidents recorded on official status pages (minor latency and unresolved ongoing items are excluded). Uptime is our own estimate based on that, not official vendor figures. Aggregated by calendar month in Japan time (JST).
Highlights of the month
September 2026 saw 31 major incidents totalling 10,036 minutes across 7 providers. Cursor logged the most (9), followed by Anthropic (8), OpenAI (5) and GitHub (4).
Nearly 60% of the total (5,961 minutes) came from a single Anthropic incident: degraded functionality for Claude Cowork on Windows. It ran for about four days, from September 11 00:54 to September 15 04:14 (JST), but only affected Cowork on Windows — claude.ai and the API were not involved. Excluding it, the month totals 4,075 minutes, almost the same as August (4,153).
The most severe incident was OpenAI’s “Issues with Codex” on September 26, 07:58–08:54 JST. Codex Web, the CLI, the VS Code extension and the Codex API were all in full outage (the highest severity) for about 56 minutes. Anthropic also saw elevated errors across claude.ai, Claude Code, Cowork and the API for about two hours from September 29 23:21.
Six of Cursor’s nine incidents involved its new Grok Bot feature (including about 6.5 hours from 23:27 on September 16 and two incidents from the evening of the 18th into the morning of the 19th — roughly 19 hours in total). New features drawing the most failures was a pattern at OpenAI too: ChatGPT Work logged three major incidents (including about 5h24m from September 1 00:04) plus at least five minor degradations.
On the developer side, GitHub recorded about 10.5 hours of “degradation with Gemini 3.8 Flash” from September 16 16:20 — a case where an upstream model provider’s trouble surfaced directly as a GitHub (Copilot) incident. September 13 brought an incident spanning several GitHub services (about 1.5 hours), and Copilot code review stopped working twice.
In image generation, Stability AI returned 500 errors on its generation API for about four hours from September 15 01:33.
Trends & takeaways
Upstream-to-downstream cascades were visible again. On September 3 at 22:26 JST, Anthropic announced elevated errors across multiple models; 15 minutes later Cursor declared an incident at its highest severity, and 51 minutes later it separately announced elevated errors for Anthropic models. The same evening OpenAI also recorded elevated errors in ChatGPT and Codex (minor).
Freshly launched features drove the counts at every company. Cursor’s Grok Bot and OpenAI’s ChatGPT Work both failed repeatedly in short bursts. If you build a new feature into your workflow, keeping a fallback ready for the first weeks is a safe bet.
Minor degradations (latency-level) numbered 72, led by OpenAI (24), Cursor (22) and GitHub (10). OpenAI had only five major incidents but announced many narrow ones — region-specific errors in Europe and APAC and plan-specific error spikes.
Anthropic shows the lowest uptime at 84.93%, but that is almost entirely the long Windows Cowork degradation above. Excluding that one incident, Anthropic’s uptime is about 98.7%. A narrow-scope issue that the provider keeps listed as a partial outage for days weighs heavily in the numbers — worth discounting when comparing.
How to read this report
Incident counts and downtime include only incidents on each provider’s official status page that were both “major or worse (effectively unusable)” and already resolved (calendar month, Japan time). Minor degradations such as latency, and unresolved ongoing items, are excluded from the totals so the figures aren’t inflated beyond real impact.
Two OpenAI items were still unresolved at month end (support chat unavailable, and elevated errors across ChatGPT, Codex and the API since September 30), but both were minor or lower in severity and do not affect the totals.
About 60% of this month’s total is the single Windows Cowork degradation at Anthropic. Subtracting it gives a more realistic picture of the month overall.
Uptime is our own estimate based on that major-outage time — not official SLA figures.
Each service’s “outage history” page lists individual incidents with title, severity, duration and the official link.