The reflex, when a model starts returning errors, is to doubt your own machine and switch browsers or sign out and back in [20]. The answer is usually published: OpenAI at status.openai.com, Anthropic at status.claude.com, xAI at status.x.ai, with Gemini in the Google Workspace Status Dashboard [21]. Those pages lag, though. On September 3 the Downdetector reports spiked before the official entries appeared, with ChatGPT peaking above 35,000 in the US [19] [7].
Ninety days is 129,600 minutes [1]. OpenAI's 99.64 percent leaves 467 of them unavailable, just under eight hours, and Anthropic's 99.4 percent for claude.ai leaves 778, a little over thirteen [11] [12]. The figure sitting beside OpenAI's changes how you read it: the APIs run at 99.94 percent and Codex at 100 [8]. Over the same window 99.94 percent is about 78 minutes, so the consumer app was out roughly six times as long as the interface developers build against [2] [3].
That gap matters less than what the afternoon of September 3 did to the arithmetic of diversification. If ChatGPT and claude.ai failed independently, the joint probability would be 0.36 percent times 0.6 percent, which over 90 days works out to 2.8 minutes of simultaneous downtime [4]. OpenAI's incident ran 14:58 to 16:55 UTC and Anthropic's second ran 13:26 to 16:16 [2] [4], an overlap of 78 minutes [5], about 28 times the quarter's independent allowance in a single afternoon [6]. Anthropic logged a shorter Claude Sonnet 5 incident from 12:37 to 12:56 UTC as well [3]. To keep this honest: "elevated errors" is degradation rather than a blackout, and OpenAI logged 60 incidents between July 1 and September 3, roughly one a day [9], almost none of which anyone outside noticed.
Neither company has given a cause in its incident notes [16]. Much of the coverage pointed at Azure East US, where ChatGPT, Claude and Grok are said to draw compute [14], but Microsoft's Azure status history carries no East US post-incident review for that day, and its most recent entry there is July 23, for West US [15]. Anthropic was willing to name "an issue with an upstream cloud provider" for an August 28 problem affecting Claude Cowork and Claude Code on the web [17], so it attributes when it chooses to. Gemini was the service that kept running while the other three were degraded, and it drew around 500 Downdetector reports without ever posting an outage notice [23] [6]. On the record, that reads as luck rather than any documented difference in architecture.
The exercise worth doing is not a percentage. For each place a model sits in the path of an answer somebody is waiting for, write down the longest continuous outage that workflow can absorb, then hold it against the two durations already on the record: 117 minutes for OpenAI's September 3 incident, 170 for Anthropic's second [7] [8]. Anything that can wait three hours needs a queue and an honest message to the user, not a second vendor. Anything that cannot wait twenty minutes needs a fallback that does not run through the same afternoon, and the option Notebookcheck will vouch for is a model on hardware you control with no internet connection [24]. There is also a cost that rarely makes it into anyone's budget: OpenAI told Codex users on the mobile remote control they may need to pair their device again after the incident [18], which is how a resolved status page becomes another twenty minutes of somebody's Friday.
