OpenAI Astra: Ten Math Results, Lean Proofs & Compute Cost

Ten problems, one of them a full solution to a 1999 open question

OpenAI attributed ten mathematics and theoretical computer science results to Astra, an internal model the company had not released, in a post published August 1, 2026. The results span high dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography and extremal combinatorics. The one full construction in the batch establishes the existence of a non sofic group, a question that had been open since 1999. [2][3]

The other named result described as a full resolution rather than an improved bound is a disproof of Connes's rigidity conjecture. The Decoder also reported new bounds for sphere packing among the ten, and both outlets noted problems posed by Paul Erdos among the set, consistent with the summary's count of three Erdos problems resolved. Most of the remaining results are improved bounds on existing problems rather than new full solutions. [2]

Atlas interpretation: Framing nine of the ten as bound improvements rather than full solutions matters for how the claim should be read. A tightened bound is real progress and can be the more useful result mathematically, but it is a different kind of claim than resolving whether an object exists at all, which is what the non sofic group construction and the Connes disproof do. OpenAI's own post keeps that distinction, and Manchester mathematician Thomas Bloom's comment to The Decoder, that the construction work is bigger in significance than a proof of the unit distance problem would have been, reads as a judgment about that same distinction rather than about the batch as a whole. [2]

Lean certificates and a small compute bill

Each write up shipped with a machine checked proof certificate in the Lean formal verification language, published on GitHub. Human researchers worked with Astra to prepare the mathematical arguments as manuscripts, and Astra then formalized each one into its Lean certificate, a division of labor that lets Lean's checker rather than a human referee confirm the logic holds. [3][2]

OpenAI stated that the total tokens needed to find solutions to the ten problems would cost roughly $2,000 at API rates for the model it calls Sol internally. OpenAI researcher Noam Brown added, in a remark BleepingComputer quoted, that the batch included no Millennium Prize Problems yet. [3]

Atlas interpretation: The dollar figure is doing rhetorical work distinct from the mathematics. A few thousand dollars of inference is a small number next to the years of unpaid effort that typically go into an open problem, and OpenAI is using it to argue that the bottleneck on this kind of research is shifting from human attention to compute budget. That argument depends on Astra's outputs holding up under the field's normal scrutiny once full papers circulate, which as of the coverage on the timeline had not yet happened. [3]

A model announced by its results, not by a release

Astra remained unreleased at the time of this post. Both The Decoder and BleepingComputer described the mathematics results as OpenAI's way of introducing Astra, which the companies characterized as OpenAI's next major model or model family, ahead of any public launch or announced release date. [2][3]

Atlas interpretation: Leading with math proofs rather than a benchmark table or a chat demo is a specific choice of evidence. Benchmark scores are easy for a reader to distrust and hard to independently rerun; a Lean certificate on GitHub is checkable by anyone with the toolchain, which shifts the claim from OpenAI's word to a verifiable artifact. That is a stronger form of proof than most model announcements offer, even though it says nothing about how Astra performs on the far larger set of tasks that are not machine checkable proofs. [2]