OpenAI's New Model Astra Solves 10 Fields Medal-Level Math Problems
0xBroomberg
OpenAI published a 249-page paper revealing that its unreleased model Astra achieved major breakthroughs on 10 long-standing open math problems — at a total compute cost under $2,000. This means AI is now pushing the frontiers of pure mathematics at the price of a graduate student's weekend stipend.
How hard are these ten breakthroughs?
Astra's results span high-dimensional geometry, coding theory, group theory — the mathematics of symmetry — operator algebras, quantum complexity, and more: at least seven distinct subfields.
Under Epoch AI's OpenMath scoring framework, most results rank as "Major Advance"; one earned a "Breakthrough" rating, potentially the top math result of the year.
This means → Astra did not drill deep in one direction. It advanced multiple unrelated mathematical frontiers simultaneously — the breadth itself is the signal.
A 27-year-old open problem — how did AI solve it?
The headline result: Astra constructed the first-ever non-sofic group, disproving a conjecture posed in 1999 by Abel Prize laureate Mikhail Gromov. For 27 years, no top mathematician managed to build a counterexample.
In plain terms = Gromov conjectured that every countable group is sofic — a type of group that can be approximated by finite symmetric structures. The world tried for 27 years and found nothing. Astra built one directly.
Astra's method: it extracted the unit group of a binary Leavitt algebra from a mathematical code library, combined Kun–Thom extension-graph theory with Thompson's group V, and delivered a complete proof — formally verified in Lean 4, a proof-assistant language, with a machine-checkable certificate.
A Caltech math PhD called this a "Fields Medal–level result." Mathematician Elliot Glazer confirmed it as "the most important AI-assisted math result to date."
A boundary frozen for 46 years — how far did it move?
The high-dimensional sphere-packing problem — how densely you can pack spheres in higher dimensions: since two Soviet mathematicians established the limit in 1978, 46 years passed with no one advancing it by even a few decimal places.
Astra delivered a new proof and computed the exact exponential decay rate of the Cohn–Elkies linear program, breaking the 1978 boundary for the first time.
This means → this was not an incremental tweak to an old method. Astra found an entirely new path through a decades-long deadlock.
A Fields Medalist's conjecture — overturned?
In 1982, Fields Medalist Alain Connes — founder of noncommutative geometry — proposed the "rigidity conjecture": that a certain class of groups generates von Neumann algebras — mathematical structures describing quantum-mechanical operators — as unique as fingerprints.
Astra proved the conjecture false, then went further: it constructed a countably infinite family of groups that are pairwise non-isomorphic yet generate exactly the same von Neumann algebra.
In plain terms = Connes believed each group's "fingerprint" could never repeat. Astra produced infinitely many groups with identical fingerprints but completely different identities — a maximal disproof.
$200 per problem — what does that cost mean?
Total cost for all ten proofs at Sol API pricing: under $2,000. That is roughly $200 per problem — about a graduate student's weekend stipend.
This means → problems that once required top mathematicians working for years or decades now carry a marginal cost comparable to a business dinner.
OpenAI researcher Noam Brown said the team has not yet cracked the Riemann hypothesis or other Millennium Prize problems, but test-time compute is far from maxed out — and "million-dollar-class" open problems "could plausibly be solved."
What is the real takeaway here?
These ten results were accidental byproducts of OpenAI's internal evaluation of Astra — not a dedicated research campaign. The model itself remains unreleased.
This reflects something larger: frontier models may already possess mathematical reasoning capabilities far beyond public perception — capabilities that simply have not been systematically unleashed.
"AI godfather" Geoffrey Hinton previously predicted that AI could produce mathematics humans cannot understand within ten to twenty years. This 249-page paper may be an early footnote to that prediction.
Content is for reference only, not financial advice.