AI Outage History
What has actually gone wrong, and for how long. Each page below lists the incidents a provider has published for one tool, with dates, severity and duration.
Why incident counts are a poor measure of reliability
Providers differ enormously in how openly they report. OpenAI posts short notices for problems many companies would never mention, so its list is long. Others publish only substantial outages, so their lists look clean. Comparing row counts across these pages measures reporting habits, not reliability.
Duration and severity are the numbers worth reading. A month with ten brief minor incidents is a better month than one with a single four-hour outage, and only major and critical incidents count against the uptime figures shown on each tool's status page.
What these lists cannot tell you
Two things are missing by design. The first is slowness: a service can be entirely available and still unusable because demand has outgrown capacity. Queued video generations on Sora or Runway are the clearest example, and they never appear as downtime.
The second is dependency failure. Cursor and GitHub Copilot both rely on models from other companies, so an incident at Anthropic or OpenAI breaks them while their own records stay clean. A quiet history alongside a bad week of your own is usually this.
Coverage also varies. Providers cap how far their public feed reaches, so an empty stretch further back means the record has aged out rather than that nothing happened.
🔗 Related