This New AI's Biggest Change Is Not Doing Your Current Job Faster

As early reactions to OpenAI's GPT-6 Astra came in, Nathaniel Whittemore, a writer and analyst who covers artificial intelligence, said the mixed response made sense. He described the model as "both significant and confounding." It was built from an entirely new foundation training, the long initial training that teaches a model how to behave, and intended to compete with Anthropic's Fable and Mythos models. He found the weekend demos, such as games made from a single request and three-dimensional designs, "really impressive, at least visually."

His line, as Whittemore put it, was this: "GPT-6 Astra is not an efficiency AI model." A minute later he added, "It is not about doing what you currently do better. It's about expanding what you can do."

What he meant, he argued, was the difference between getting faster at familiar work and getting access to unfamiliar work. Earlier systems like GPT-3.5 and GPT-4, he said, felt instantly important because they landed on work many people already did, looking up answers, researching, and writing, through a chat interface people already knew, trading typed messages with a chatbot.

In his account, Astra's advance is running software itself. A person talks while the system carries out the steps inside programs and web pages. OpenAI's chosen numbers fit that emphasis, Whittemore noted. Astra edged Fable 5.1 on the coding test Terminal Bench 4.0, 57.6 percent to 55.8 percent, at much lower cost, and pulled far ahead on Automation Bench, which tests hands-on software operation, 41.1 percent to 31.4 percent, which OpenAI's announcement calls "the world's best computer use model."

He argues the changes behind the move to computer use "are immense, but not easy to package up in a weekend of testing." In his argument, the model looks puzzling when judged only on familiar office tasks, because, in his view, its value is in opening work people could not do before.