Line up the three windows and the morning reads as a relay. Anthropic's incident opened at 7:37 a.m. ET and closed at 11:27 [5]; the four-hour Claude span the CIO account summarises [2] computes to 3 hours 50 minutes from the timestamps the same article prints [4]. Grok ran from about 9:30 a.m. to 1:08 p.m. [7]. OpenAI's sat in the tail, about 11 a.m. to 12:55 p.m. [3]. At least one of the three was impaired for 5 hours 31 minutes [2]. All three were impaired at once for 27 minutes, the stretch between OpenAI's start and Anthropic's fix [1]. For the remaining 5 hours 4 minutes, a client that failed over on error rate had somewhere to go [3].
Two of the three incidents took in API tiers and developer tooling [4][7], so this reached scheduled and automated work rather than only chat sessions. Anthropic's was a list of models returning elevated errors rather than an endpoint refusing connections [5], and the same signature had produced a roughly 27-minute Sonnet 5 incident the previous day, a coincidence of duration and nothing more [6]. Degraded-but-answering is the harder case to route around, because a retry loop against elevated errors spends its budget while continuing to look like it might succeed.
Cause is open. The common-dependency theory offered by Info-Tech Research Group's Brian Jackson, if it were confirmed, would make brand diversity and dependency diversity two different purchases, and enterprises have mostly been making the first [8]. Sitting next to it is a confound: OpenAI's window fell on the day the company launched GPT-6 Astra, the frontier model it says approximates AGI [9]. Nothing in the account ties the three incidents to a single root cause [14], and a shared-layer explanation has to account for the launch landing inside the same hours.
The thing the day does not measure is exposure. There is no count of stalled workflows and no share of enterprises running multi-provider failover [15]. Technology analyst Carmi Levy's judgment that firms could be "uncomfortably exposed" and that the risk is no longer hypothetical [10] is a fair account of the direction of travel, though the baseline he offers for it (that a few months ago enterprise AI was mostly chatbot answers and simple email drafts [11]) is recollection rather than measurement. Whether anyone actually had a second provider wired into an agent path went unreported.
So the figure worth keeping is 27 minutes. That is the window no amount of provider switching covers, and it is short enough that a queue with a retry ceiling and a notified human absorbs it. The five hours either side are a routing problem, and routing problems reward whoever wrote the fallback before the morning it was needed.
