ChatGPT, Claude and Grok went down at once: companies need a contingency plan for AI outages

On Thursday, almost simultaneously and without a clear explanation, ChatGPT (OpenAI), Claude (Anthropic) and Grok (SpaceXAI) suffered prolonged outages that affected users and the enterprise workflows that now depend on these systems. The episode, originally covered by Computerworld in a piece by journalist Taryn Plumb, has raised an uncomfortable question for technology leaders: what happens when agentic AI assistants, increasingly embedded in critical processes, simply stop responding?

The three services went down hours apart but within the same morning window, US Eastern time. ChatGPT was affected at around 11:00 a.m. ET for roughly two hours, with problems in search, file uploads, agents, GPTs, voice mode, image generation, ChatGPT work, the Compliance API, Deep Research, ChatGPT Atlas and other connectors and applications; Codex services (web, API, command line and the VS Code extension) were also affected. OpenAI resolved the incident at around 12:55 p.m. ET and advised Codex remote-control users to re-link their mobile devices. Notably, the outage coincided with the scheduled launch of GPT-6 Astra, the new frontier model that, according to OpenAI, approaches artificial general intelligence (AGI).

Claude began failing earlier, at around 7:37 a.m. ET, with elevated error rates across an extensive list of models — Mythos and Fable 5.1 and 5, Sonnet 5, and Opus 5, 4.8 and 4.6 — until it was resolved at around 11:27 a.m. ET, some four hours later. Anthropic had already suffered a shorter interruption the previous day, lasting about 27 minutes, also due to elevated errors in Sonnet 5. Grok, for its part, began showing problems at around 9:30 a.m. ET, affecting Grok Web, Build, API, the Office/Workspace plugins, Android and X, until traffic returned to "healthy" levels at 1:08 p.m. ET, close to three and a half hours later.

Brian Jackson, principal research director at Info-Tech Research Group, called it "curious" that several different providers failed at the same time, and floated the hypothesis that they might share some common infrastructure element, such as a CDN layer, DNS or shared cloud infrastructure. That is a theory, not a confirmed cause: the coincidence itself is described as largely mysterious.

Technology analyst Carmi Levy went further in assessing the incident's implications, arguing that it should serve as a "wake-up call" for IT leaders who have largely ignored the cost of these increasingly critical platforms suddenly going dark, because "the risk is no longer hypothetical." According to Levy, only a few months ago AI use at the typical company was limited to chatbots answering basic questions or drafting simple emails, so large-scale outages barely dented overall productivity. That is changing fast: organizations now need a better understanding of the impact agentic AI has on day-to-day work and of the extent to which it disrupts employees' ability to complete complex tasks once they have handed control to automated cloud tools.

Levy also warned of a side effect: when the tools go down, employees can fall back on traditional manual workflows — updating spreadsheets or producing reports by hand — but they may also discover that, after relying so heavily on AI agents, their cognitive skills "may not be as sharp as they used to be." He therefore recommends revisiting disaster recovery and business continuity plans, documenting workflows in greater detail, designing recovery scenarios for a prolonged outage and stepping up training so staff maintain their manual skills over time. Unlike productivity platforms such as Google Workspace or Microsoft 365, which offer some "offline mode" with locally stored data or syncing with Dropbox or Google Docs for desktop, agentic AI platforms currently offer far fewer alternatives for continuing to operate offline. The more companies delegate critical tasks to agents — and the further they push people out of the process in the name of productivity — the less prepared employees will be to take back control when the service goes down, something Levy considers inevitable.

Jackson, from Info-Tech, proposes a more architectural way out: treating the language model as a "commodity" that can be hot-swapped for an alternative, whether another cloud provider (which, with luck, is not suffering a simultaneous outage) or a self-hosted option such as an open-weights model. The idea is to have a backup layer of "intelligence" available when the main provider is not, even if only as a stopgap.

Taken together, the episode leaves a conclusion both experts share: many organizations, in their rush to automate processes with AI, have not put the same effort into planning what to do when that AI fails. As Levy sums it up, "it is entirely possible for well-intentioned organizations to over-rely on AI automation," and "too many are about to learn hard lessons for not having a contingency plan."