OpenAI’s Astra Claims Ten Major Mathematics Advances

OpenAI has presented ten claimed advances on long-standing open problems in mathematics and theoretical computer science. The company says it deliberately selected questions whose principal result had not progressed for at least a decade, with most having remained unresolved for considerably longer. The collection reaches across high-dimensional geometry, error-correcting codes, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. These are distinct specialties, but they share a central feature: each asks for a rigorous argument that must survive detailed mathematical scrutiny rather than merely produce a plausible numerical answer or an experimentally successful prediction.

The work is attributed to an internal version of Astra, which OpenAI describes as its next major model. According to the company, the computational search that found the solutions consumed tokens whose total cost would be about $2,000 at Sol API prices. Humans then worked with the same model to turn the arguments into manuscripts. The model also formalized every argument in Lean, a proof assistant that represents definitions and logical steps in a machine-checkable form. If the results withstand independent expert review, the important leap is not simply that an AI answered difficult questions, but that one system reportedly generated arguments across many mathematical fields and converted them into formally verifiable certificates. OpenAI is additionally publishing a narration of the model’s reasoning process for each proposed solution, giving researchers material with which to inspect how the arguments were developed.

Two results concern packing and coding in high-dimensional spaces. For sphere packing, OpenAI reports new upper bounds on how densely non-overlapping spheres can be arranged as dimension grows, reaching the Cohn–Elkies threshold. An upper bound does not construct the densest packing; instead, it proves that no possible arrangement can exceed a stated density. In coding theory, the company claims exponentially improved bounds for binary codes, along with analogous conclusions for high-dimensional spherical codes. Codes are collections of well-separated representations used to distinguish messages despite errors, while spherical codes arrange separated points on a high-dimensional sphere. Exponential improvements are especially notable because their advantage grows rapidly with the relevant dimension or size parameter rather than providing only a fixed numerical refinement.

In algebra and operator theory, Astra reportedly produced a construction proving the existence of non-sofic groups and an argument disproving Connes’ rigidity conjecture. Group theory studies abstract systems of symmetry and composition, while the sofic condition identifies groups that can, in a broad sense, be approximated by finite permutation structures. Establishing that non-sofic groups exist would therefore settle an existence question by showing that this approximation framework does not cover every group. The claimed result on Connes’ rigidity conjecture has a different logical character: rather than establishing a conjectured universal principle, it supplies a disproof. Both cases illustrate why explicit reasoning matters. A convincing resolution must expose a construction or counterargument that specialists can examine line by line.

The arithmetic circuit result addresses lower bounds for computing the permanent. The permanent resembles the determinant of a matrix but lacks the determinant’s alternating signs, and computing it is a central hard problem in complexity theory. An arithmetic circuit models a computation assembled from algebraic operations. Proving lower bounds means showing that circuits of a specified kind cannot compute the permanent unless they have sufficient size or complexity. Such lower-bound questions are notoriously difficult because researchers must rule out whole classes of possible algorithms, not merely demonstrate that known methods are inefficient. OpenAI characterizes its contribution as new lower bounds rather than a complete resolution of every broader complexity question surrounding the permanent.

For quantum complexity, the company reports an exponential parallel repetition theorem for general two-player quantum games. Parallel repetition asks what happens when an interactive game is repeated multiple times: ideally, the probability of winning all repetitions should fall rapidly unless the players could already win the original game with certainty. Quantum strategies make this harder to analyze because players may share entanglement. The claimed theorem gives exponential behavior for the general two-player quantum setting, according to OpenAI. In lattice cryptography, Astra is said to have established polynomial-factor hardness of approximation for the closest vector problem. That problem asks for a lattice point nearest to a target, and approximation hardness measures how difficult the task remains when an algorithm is allowed to return a point that is only within some factor of optimal. The announcement links the result to a problem area that underlies important cryptographic research, without claiming a specific immediate change to deployed systems.

The remaining advances concern discrete geometry and extremal combinatorics. OpenAI says Astra determined the maximum possible volume in Ehrhart’s volume conjecture. It also reports a superexponential lower bound for multicolor triangle Ramsey numbers, resolving Erdős problem 183. Ramsey theory studies the unavoidable patterns that appear once a structure becomes sufficiently large; in the multicolor triangle setting, the question concerns when a monochromatic triangle must occur under edge colorings. A superexponential lower bound shows that the threshold grows even faster than an ordinary exponential scale in the relevant parameter. Finally, the company claims results on compactness and degeneracy conjectures in extremal graph theory, resolving Erdős problems 146 and 180. Extremal graph theory asks how large or dense a graph can be while avoiding specified configurations, so these resolutions concern structural limits rather than a single computed instance.

The breadth of the announcement is central to its significance: the ten claims are not variations on one benchmark but proposed contributions to separate research communities, including several questions associated with major mathematical conjectures and Erdős problem lists. At the same time, the publication should be understood as the beginning of community evaluation, not as a substitute for it. Lean certificates can check whether a formal proof follows from its encoded assumptions and definitions, but mathematicians still need to inspect whether those formal statements faithfully represent the intended problems, assess the manuscripts, and place each result in its existing literature. OpenAI says it accepts responsibility for correctness, which makes the quality and transparency of the released manuscripts, certificates, and reasoning records especially consequential.

OpenAI presents the project as part of a wider effort to build tools that accelerate scientific and mathematical discovery. It connects these results to an earlier AI-generated disproof of the Erdős unit-distance conjecture and to ChatGPT for Academic Researchers, an initiative offering 100,000 scientists and mathematicians free access to advanced ChatGPT models. Together, those efforts suggest a strategy that combines direct attempts at original research with broader access to AI assistance. The stated token cost is also striking: although it excludes the full human, training, infrastructure, and validation costs behind the system, it indicates that the marginal inference used to search for these particular solutions was comparatively modest at the cited API rates.

OpenAI explicitly acknowledges that research-capable AI creates questions that a technology company cannot settle alone. Attribution is one of them. The company says human contributors helped prepare the manuscripts and formalize the proofs with the model, but that the mathematical arguments themselves were generated by Astra. This distinction matters because conventional authorship practices were built around human intellectual contribution and responsibility. By openly assigning the generation of the arguments to the AI while accepting responsibility for their correctness, OpenAI is forcing a practical debate about credit, accountability, verification, and authorship in AI-assisted science. The lasting importance of the release will depend both on whether specialists validate the ten results and on whether its evidence makes AI-generated research reproducible and trustworthy enough to become part of ordinary mathematical practice.