OpenAI’s unreleased Astra model has solved ten long-standing open problems in mathematics and theoretical computer science, each unresolved for at least a decade, and published machine-checkable proofs that any researcher can independently verify. The company released a 249-page manuscript alongside Lean 4 proof certificates on GitHub, with the total compute cost estimated at roughly $2,000 at Sol API rates. The results demonstrate that frontier AI systems can now generate original, formally verified contributions to pure mathematics.
The ten problems resolved were announced August 1 and span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. The headline result is the first explicit construction of a non-sofic group, answering a question that stood open since Mikhail Gromov introduced soficity in 1999. Soficity relates to whether every group can be approximated by finite groups, a property that underpins results in operator algebra theory, ergodic theory, and the study of cellular automata. Astra’s construction settled the question definitively: not every group is sofic.
Other breakthroughs include new upper bounds on high-dimensional sphere packing density down to the Cohn-Elkies threshold, exponentially improved bounds on binary and spherical codes, a disproof of Connes’s rigidity conjecture on group von Neumann algebras, an arithmetic-formula lower bound for computing the permanent, and a superexponential lower bound for multicolor Ramsey numbers resolving ErdΕs problem 183. These are not incremental refinements. They are new constructions, counterexamples, and bounds that professional mathematicians had not produced.
Machine-Verified Proofs Change the Game
OpenAI released Lean 4 proof certificates under an Apache 2.0 license, and the repository shows zero unresolved steps. A Lean proof that type-checks is valid by construction. No trust in OpenAI’s claims or benchmarks is required. Any mathematician or independent researcher can pull the certificates from the openai/ten-proofs repository and verify every step directly. That shifts the relevant question from whether the proofs are correct to whether the mathematical ideas they contain are genuinely new and how they were found.
The compute cost is striking. The total token cost for all ten solutions was roughly $2,000 at GPT-5.6 Sol API rates, according to OpenAI’s own estimate. A mathematical research budget that any funded university group could authorize is a fundamentally different scale from the millions of dollars that characterized earlier frontier AI research. As inference costs continue to drop, the constraint on mathematical discovery may increasingly shift from human talent scarcity to compute availability and problem selection. This trajectory is consistent with the broader AI market’s rapid growth and declining inference costs across the industry.
Fields Medalist Timothy Gowers said he would recommend one of the proofs for publication in the Annals of Mathematics without hesitation. Thomas Bloom, who maintains the erdosproblems.com database, called the results big news and said they were bigger as mathematical constructions than the earlier ErdΕs unit-distance counterexample. But the broader mathematical community is still evaluating the manuscripts. Full peer review has not been completed, and the community is divided on whether AI-generated research should follow the same publication pathways as human work.
Tensions With the Mathematical Community
The announcement arrived amid broader tensions. In June, mathematicians endorsed the Leiden Declaration, warning that AI companies are announcing results through press releases rather than peer-reviewed journals, using published research without consent, and threatening the integrity of proof and attribution. The declaration was endorsed by the International Mathematical Union. OpenAI acknowledged the concerns and said it was engaging with the mathematical community. On August 17, the company paused Astra’s broader deployment over safety concerns, though the math results themselves remain published and verifiable on GitHub.
OpenAI credits Astra with the underlying mathematical reasoning while human researchers prepared the arguments into publication-ready form. The company said it intends for Astra to become its primary research tool, with a public release targeted within a year. The model would need to pass a government security review under the new TRAINS framework before any public launch. The TRAINS process, which reviews advanced AI systems for national security risks, represents a new regulatory layer between research breakthroughs and public access.
The results also arrived just as the European Union’s AI Act reached full enforcement on August 2, 2026, creating new obligations for companies deploying high-risk AI systems. While the math proofs themselves fall outside the Act’s direct scope, the regulatory environment reflects growing government scrutiny of AI capabilities.
For the mathematical community, the central question is whether AI-generated mathematics will be judged by the quality of its proofs or the process that produced them. If a machine can formalize a proof that stumped mathematicians for decades and attach a machine-checkable certificate, the traditional publication pipeline may need to adapt to a world where new knowledge arrives with formal verification built in. For researchers, the most immediately actionable element is the publicly available proof repository, which provides a template for how AI-assisted mathematical research can be verified independently of any lab’s claims.
discussion