OpenAI gave U.S. policymakers and regulators an early look this week at a new family of AI models tentatively named "Astra," according to a report from The Information, surfaced by Techmeme.
The pitch, per The Information, centered on one capability in particular: the models' improved ability to complete long-running tasks. That's the industry's shorthand for work that unfolds over many steps and stretches of time — the difference between a chatbot answering a question and a system that can grind through a multi-stage job without a human nudging it along at every turn.
The Information reports that OpenAI is preparing to release the model family, though the "Astra" name is described as tentative and could change before launch. Beyond the demonstration itself and the long-task claim, few details have been reported — no release date, no pricing, no benchmark figures.
What stands out as much as the technology is the audience. Rather than debuting the models at a developer event or in a public blog post, OpenAI took them to Washington first, showing them to the officials who write and enforce the rules that AI companies operate under.
That sequencing has become a familiar pattern for frontier AI labs, which have an interest in shaping how regulators understand a new capability before the public forms its own impression of it. Neither source reports how the officials responded.
It matters because models that can carry out long, multi-step tasks on their own are the closest thing yet to AI that does jobs rather than answers questions — and Washington is being shown that shift before the rest of us get to try it.