The Cheapest Breakthrough in Math: OpenAI's Astra Delivers Ten New Proofs for $2,000
- David Borish
- 5 days ago
- 5 min read
On August 1, 2026, OpenAI released a 249-page manuscript containing ten new results in mathematics and theoretical computer science, each attributed to an internal version of an unreleased model the company calls Astra. Every result addresses a problem that had seen no progress on its main statement for at least a decade, and in most cases far longer. The company published Lean 4 formalizations for all ten on GitHub, along with chain-of-thought walkthroughs of the model's reasoning. OpenAI states that the tokens used to generate all ten arguments would have cost about $2,000 at its Sol API rates.
Two things separate this release from the steady stream of "AI does math" announcements. The first is the verification method. The second is what the results are actually about.
What Lean certificates change
A Lean certificate is not a summary or a confidence score. Lean is a proof assistant with a formal kernel, and every logical step in a submitted argument has to check out against that kernel or the proof does not compile. Publishing the .lean files means an outside mathematician does not have to trust OpenAI's prose description of what the model did. They can run the verifier and read the compiler output.
That distinction matters because language models have a documented history of producing confident, fluent, and wrong mathematics. A narrative claim that a model "solved" a conjecture invites the reasonable response of asking who checked it. Machine-checkable certificates move part of that burden onto a compiler. The kernel does not care about fluency, and it does not accept hand-waving between steps.
The certificate covers the logical validity of the formalized statement. It does not, by itself, settle two other questions that mathematicians will raise: whether the formalized theorem matches the informal conjecture as the community understands it, and how much human problem-shaping preceded each model run. OpenAI addresses attribution directly in the paper. The company says it prepared the manuscripts and formalized the Lean proofs, takes responsibility for their correctness, and credits the mathematical arguments themselves to the model. It also cites the Leiden Declaration on AI and Mathematics, a June statement endorsed by the International Mathematical Union that criticized AI companies for announcing results through press releases rather than peer review, and it asks the mathematical community to engage with the results rather than accept them.
The results
The ten problems span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. A few are worth describing concretely, because the abstractions hide how specific the claims are.
The headline result is the construction of a non-sofic group. Soficity is a property Mikhail Gromov introduced in 1999, describing groups that admit finite permutation approximations. Whether every countable group has this property had been open for 27 years. The manuscript constructs an explicit group that does not, using property-(T) expanders and the binary Leavitt algebra. This resolves the question in the negative: not every group admits such approximations.
The sphere-packing result determines the exact exponential decay rate of the Cohn–Elkies linear program, the Fourier-analytic method that Maryna Viazovska used to settle the optimal packing in dimension eight. The paper proves that the program's optimal density bound decays at rate the square root of e over 2π, which works out to a packing exponent improvement. The classical Kabatianskii–Levenshtein exponent from 1978 was 0.59905576. The new bound gives 0.6044, the first improvement to the general high-dimensional sphere-packing exponent since 1978. The paper also proves a matching lower bound, showing that no Cohn–Elkies auxiliary function can do better, which closes the question rather than nudging it.
A companion result improves classical upper bounds for binary and spherical codes by exponential factors across all parameters, the first improvements to those general high-dimensional exponents since 1977 and 1978 respectively. In the small-distance limit, the spherical construction recovers the same sphere-packing exponent obtained by the Cohn–Elkies analysis, so two independent lines of argument arrive at the same number.
The remaining results include a disproof of Connes's rigidity conjecture on von Neumann algebras, a proof of Ehrhart's volume conjecture in every dimension, new circuit-complexity lower bounds for computing the permanent, a parallel repetition theorem for two-player quantum games, hardness results for the Euclidean closest vector problem relevant to lattice cryptography, a superexponential lower bound on multicolor Ramsey numbers, and separate constructions disproving two conjectures in extremal graph theory attributed to Erdős and Simonovits. Several of these correspond to numbered problems in Erdős's catalogue, including problem 183 on multicolored Ramsey numbers.
The price is the argument
OpenAI's $2,000 figure is for all ten proofs combined, not per problem. Whatever weight the individual results carry once specialists work through them, that number is the part that reframes the discussion. Sphere-packing exponents and non-sofic groups are the kind of problems that consume years of a specialist's career. If the arguments hold, the cost of producing a research-level result on a decade-old problem drops to a rounding error in a lab's compute budget.
That is the transfer worth watching. Mathematical reasoning proven out in a controlled setting, with a compiler as the referee, becomes an input that scales with compute rather than with the supply of specialists. Formal verification is exactly the kind of environment where a capability can be checked cleanly before anyone claims it works in the wild, which is what makes these results harder to wave away than a benchmark score.
The caveats are real and OpenAI names most of them. Astra is internal and has no release date. Nobody outside the company has run it. External mathematicians have not had time to work through arguments of this depth, which normally attract months of scrutiny. The manuscript does not fully spell out how much human framing preceded each run. A retraction on any single one of the ten would land loudly, and the community is primed to look for one.
What to watch next
The near-term test is straightforward: independent mathematicians run the Lean verifier, then argue about whether each formalized statement faithfully captures the conjecture it claims to resolve. The kernel settles logical validity quickly. The harder conversation, about faithfulness of formalization and the division of labor between model and humans, will take longer and will not have a compiler to end it.
If a majority of the ten survive that process, the more interesting question is not which lab posted the result first. It is what happens to fields adjacent to formal methods, including verification, cryptography, and optimization, once producing a checked proof on a hard open problem costs less than a plane ticket. The release does not answer that question. It makes it concrete enough to argue about with numbers.
About the author
David Borish writes long-form analysis of frontier AI research, enterprise deployment, and technology policy at davidborish.com. He is the author of the forthcoming book The Tony Hawk Paradox, which argues that capabilities first proven in controlled or simulated environments consistently transfer into broader real-world systems.
