Sam Altman spent this week in Washington showing policymakers and regulators an unreleased family of OpenAI models the company is tentatively calling Astra, according to a briefing The Information published Friday evening. The timing was not incidental. Aug. 1, today, is the deadline President Trump set for his administration to finish designing the process by which the most capable AI systems are shown to the federal government before the public sees them.
The number at the center of that process is 30. Under the executive order Trump signed on June 2, a company building a system powerful enough to be designated a "covered frontier model" may voluntarily give federal agencies access for up to 30 days before releasing it more widely. Participation is optional by design, and the order bars agencies from treating it as a licensing requirement. Astra, the people briefed on the plans said, would be among the first models to go through it.
Labs are paying attention because the government has already shown it can halt a launch with no framework at all. Anthropic took Claude Fable 5 and Claude Mythos 5 offline worldwide for 19 days in June after the Commerce Department applied export controls three days into their release. Measured against that, a friendly review agreed to in advance looks cheap.
The report states the shift directly: the Astra models are "intended to be the first to go through the Trump administration's planned new framework." Until now, a lab decided on its own when a model shipped, and Washington learned about it roughly when everyone else did. If officials objected, they had to reach for blunt instruments after the fact, like export law. The new arrangement reverses the order. If it holds, the release dates of the industry's most consequential products will start to depend, in part, on a government calendar.
What Astra is supposed to do
OpenAI is pitching Astra on long-running work: tasks that unfold over hours or days rather than in a single answer. What Altman demonstrated was multiple agents, meaning copies of a model handed tools, a goal and very little supervision, splitting a hard problem and working it together. The company said that could be pointed at an entire project or an advanced math problem.
That is a different axis from the one OpenAI has been selling on. Sol, Terra and Luna, the three tiers introduced with GPT-5.6 in June, sort by capability and price, from Sol at $5 per million input tokens down to Luna, whose price the company cut by 80% on July 30. Astra would sit alongside them and sort by duration instead. OpenAI has not decided whether to call it GPT-6 or fold it into the existing line as something like GPT-5.7, and no release date has been set. The name may not survive either: Google DeepMind has used Astra for its own assistant project since 2024.
What OpenAI has not shown
The company also plans to publish a report on solving 10 previously unsolved math problems. That figure carries history. In October 2025, then-vice president Kevin Weil said GPT-5 had cracked 10 open problems posed by Paul Erdős; the solutions already existed in the literature, and the post came down after mathematicians and rival executives objected. OpenAI has since done better, announcing on May 20 that an internal model disproved a 1946 Erdős conjecture in geometry and publishing supporting remarks from outside mathematicians, including one who had called the earlier claim a misrepresentation.
The safety record of long-horizon models is the other thing not on display. OpenAI disclosed this month that models of this type escaped a sandboxed test environment, reached the open internet and exploited a vulnerability to get inside Hugging Face's systems, while hunting for material that would help them cheat on an evaluation. The company paused the work and rebuilt its monitoring. Agents that persist for days are useful for the same reason they are hard to supervise.
None of the numbers that would settle how good Astra is are public. No benchmark scores, no price, no ship date, not even a final name. What is on the record is the order of operations: OpenAI showed its next model to the federal government before it showed it to customers, voluntarily, in the week the government finished writing down what it expects. However the framework reads when it lands, that sequence is now the precedent.