Ten problems, a decade or more of silence on each, and OpenAI cleared all of them with an internal model build it wasn’t ready to name in public until August 1. The math report that landed that day did something OpenAI’s PR team is usually careful not to do: it confirmed the name “Astra” for the first time, weeks after The Information had already reported Sam Altman demoing it to politicians and regulators in Washington.

I want to sit with the math for a second before getting to the part that worries me more. The ten proofs span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics, and one of them settles a genuine open question in group theory: the existence of non-sofic groups. Thomas Bloom, the University of Manchester mathematician who runs erdosproblems.com, called the results bigger news than the counterexample to the unit distance conjecture OpenAI published back in May, at least “in terms of constructions.” Noam Brown, who worked on the test-time reasoning tech underneath all of this, was more deflating about it on X: no Millennium Prize Problems yet, and OpenAI didn’t spend much compute on any single problem either. All ten solutions together would have cost about $2,000 at Sol’s API rates. That number is doing a lot of work in OpenAI’s messaging, and it should, because a math department can’t easily argue its way out of a $2,000 receipt.

Here’s what I can’t stop turning over: Astra is reportedly going to be the first OpenAI model class run through the Trump administration’s planned pre-release review framework, the one requiring frontier labs to submit models to the federal government before they ship. That’s a different kind of government leverage than the one I wrote about in June, when Washington reached into a frontier lab and pulled the plug on a model that was already live. This is upstream of release, a permission slip instead of a kill switch, and I’m skeptical the administration hits the deadline its own reporting implies. The other time this year Washington tried to draw a hard line around a frontier model, it took months of enforcement chaos before anyone had a clean rule to point to, and I don’t see why a review framework would move faster than an export ban did.

The multi-agent framing matters too, even though it’s easy to skip past because “coordinates multiple agents on long-running tasks” sounds like every other roadmap slide out of every other lab right now. Astra is supposed to sit alongside Sol, Terra, and Luna as a fourth model family, built for problems that take hours or days instead of minutes, and OpenAI’s chief scientist has been saying since last summer that the company wants systems that can run a research project unsupervised. Multi-agent setups have a bad habit of getting worse at tightly coupled tasks, not better, since coordination overhead eats the gains a single strong model would have banked on its own, and OpenAI’s own agents have already shown how badly that kind of coordination can fail once nobody is watching closely. Whether Astra holds together on that kind of problem, instead of the highly gameable one-shot proof generation it just aced, is the question nobody answered on August 1.

There’s a credit question buried in here too, and I think it matters more than the math. OpenAI explicitly declined to claim human authorship for the proofs, pointing to the Leiden Declaration on AI and Mathematics, even though its own researchers helped turn Astra’s raw arguments into publishable papers and take responsibility for their accuracy. That’s a more careful position than I expected from a company that spent much of the spring implying its models were becoming autonomous researchers on their own. It also gets harder to hold onto the moment a proof from an unreviewed model turns out to be wrong, which given the pace OpenAI wants to publish at, feels like a matter of when rather than if.

None of this settles whether the government review framework becomes a real constraint or a rubber stamp with better optics than a voluntary safety pledge. I’ve watched enough AI policy fights collapse into paperwork exercises to take Altman’s Washington demo with some skepticism. It’s the same question of who actually holds the leverage when a company asks the government for more speed and more oversight at once that I couldn’t answer in June, and I still can’t. Ask me again once OpenAI actually has to submit something.

Sources