OpenAI released Astra, the model it built on top of GPT-6, on September 3, giving it computer-use control (the ability to click, type and navigate a screen the way a person does, not just answer in text) and framing it as a step toward software that operates a machine on its own. A day earlier, ChatGPT, Anthropic's Claude, xAI's Grok and Google's Gemini had all gone down within the same window, an overlap Ars Technica and Wired both flagged as unusual precisely because these four run on separate model weights, separate companies and, in Anthropic's case, primarily on Amazon's Trainium chips rather than Nvidia GPUs. Separate weights did not mean separate infrastructure. When four competitors who spend billions distinguishing their models from each other go dark at the same hour, the failure sits below the model layer, in the cloud routing, the DNS, or a shared upstream vendor none of the four have named publicly.

That is the gap Astra's launch sits inside. A model that clicks and types on a user's behalf is only as reliable as the infrastructure under it, and this week that infrastructure had a bad day across four unrelated companies at once. OpenAI's own postmortem, if it publishes one, is the next document that matters here: it will either name the shared dependency (a DNS provider, a CDN, a specific Azure region) or leave procurement teams at every bank and enterprise now piloting computer-use agents to assume the outage risk is systemic, not company-specific, the next time they build an approval workflow around Astra clicking through a live production system.