For 27 years, mathematicians have wondered whether a certain kind of mathematical object — a "non-sofic group" — actually exists. It's a question so abstract that explaining why it matters takes most mathematicians a full seminar. This week, an unreleased OpenAI model called Astra answered it. Along with nine other open problems. For about $2,000 in compute.
OpenAI didn't announce Astra with a keynote or a product launch. It announced it with math — publishing ten solved problems, including formal, machine-checkable proofs that any mathematician can verify using the Lean 4 proof assistant, without having to take OpenAI's word for anything.
What Astra actually solved
The headline result concerns "sofic groups," a concept the mathematician Mikhail Gromov introduced in 1999. Roughly speaking, sofic groups are mathematical structures that can be approximated by finite pieces in a precise technical sense — and for nearly three decades, nobody could say for certain whether groups existed that couldn't be approximated this way. Astra produced the first explicit construction of one.
That wasn't the only result. Astra also disproved a well-known conjecture in operator algebra theory known as Connes's rigidity conjecture, proved a case of Ehrhart's volume conjecture in combinatorics, and resolved three separate problems from the famous open-problem catalogue assembled by the mathematician Paul Erdős, whose backlog of unsolved puzzles has occupied mathematicians for generations.
Why a Fields Medalist is vouching for it
The result that's drawn the most attention from professional mathematicians is the endorsement from Timothy Gowers, a Fields Medal winner and one of the most respected figures in the field. Gowers reviewed Astra's non-sofic groups proof and said he would recommend it for publication in a top journal without hesitation. He then joined eight other mathematicians, including Noga Alon, in writing a companion paper that translates Astra's machine-generated proof into a form other mathematicians can more easily follow and build on.
That collaboration matters as much as the result itself. It's one thing for an AI system to produce an answer; it's another for that answer to hold up when some of the most rigorous, skeptical readers in academia — professional mathematicians whose entire discipline is built on demanding proof — sign their names to it.
Verification, not trust
Part of what's made Astra's results credible so quickly is that OpenAI didn't ask anyone to simply trust the output. Every one of the ten results shipped with a formal proof certificate written in Lean 4, a proof assistant that can mechanically verify whether a mathematical argument is logically valid, line by line. Anyone with the Lean compiler installed can check the proofs themselves, without needing to trust OpenAI, the model, or any single mathematician's read of the work.
That verifiability is a big part of why the story has moved so fast through the math and AI research communities this week. Previous claims of AI-assisted mathematical breakthroughs have often required a lengthy, informal peer-review process before anyone outside the original team could be confident the result was real. Astra's results, by contrast, came bundled with their own proof of correctness.
The $2,000 figure, and why it's the real headline
As striking as the mathematics is, the number that's circulated most widely is the price tag: roughly $2,000 in API compute costs to produce ten results that, collectively, represent decades of unsolved problems across several distinct branches of mathematics. For context, funding a single graduate student or postdoctoral researcher to work on one such problem for a year typically costs tens of thousands of dollars, with no guarantee of a solution at the end.
That cost comparison is fueling a broader conversation about what "doing research" will mean going forward — not just in mathematics, but in any field where progress depends on generating and checking novel, rigorous claims. Astra hasn't been released publicly, and OpenAI has said comparatively little about the model's architecture or training beyond the results themselves. But the demonstration was deliberate: rather than making claims about capability, the company let mathematicians independently verify the work first, then talk about the model.
Whether Astra represents a one-off showcase or a genuine shift in how mathematical research gets done, the ten proofs are now public, verifiable, and — according to one of the field's most decorated living mathematicians — good enough to publish.