OpenAI's Unreleased Astra Model Solves Ten Hard Math Problems

On 1 August, OpenAI published ten results in mathematics and theoretical computer science and attributed the arguments to an internal version of a model it calls Astra. The company says each problem had "seen no progress on the main result for at least a decade," and in most cases much longer. A 249-page manuscript accompanies the announcement, along with a narration of the model's reasoning and a public repository containing a Lean 4 formalisation of every result under an Apache 2.0 licence.

The reflex with an announcement like this is to check whether "open" is doing honest work. Labs announce mathematics the way they announce benchmark scores, and the word covers both the conjectures that define a field and the questions nobody has bothered to grind through. The distinction usually collapses the claim. This list survives it.

What is on the list

The construction of a non-sofic group leads. Soficity describes how well an infinite group can be approximated by finite permutation systems. Mikhail Gromov introduced the property in 1999, and every group most mathematicians handle in practice, amenable and residually finite ones included, satisfies it. Whether every countable group must be sofic stayed unanswered for twenty-seven years. OpenAI says Astra produced an explicit counterexample, formalised in NonSoficGroup.lean.

Second is a claimed disproof of Connes's rigidity conjecture. Alain Connes posed it in 1980; Sorin Popa later sharpened it into a finite-to-one version. The question asks how much of a group survives the passage into its von Neumann algebra. According to the manuscript, Astra produced a countable family of pairwise non-isomorphic, finitely generated property-(T) groups whose group von Neumann algebras are all isomorphic. If it holds, both formulations fall.

Then: new upper bounds on high-dimensional sphere-packing density reaching the Cohn–Elkies threshold, which OpenAI describes as the first movement on the general exponent since the Kabatiansky–Levenshtein method of 1978. Exponentially stronger bounds for binary codes at every minimum distance, with corresponding results for spherical codes. An arithmetic-formula lower bound of order n⁴/log n for the permanent. An exponential parallel repetition theorem for general two-player quantum games. Polynomial-factor hardness of approximation for the closest vector problem, which sits under post-quantum lattice cryptography. Ehrhart's volume conjecture, determined in every dimension. And three entries from Erdős's catalogue: problem 183 on multicolour triangle Ramsey numbers, plus 146 and 180 in extremal graph theory.

Several of these are questions a specialist would name if asked what remains open in their subfield.

The certificates

The repository is a standard Lean 4.32.0 project built on mathlib and Lake, with one named module per result. Two commands compile all ten. Coverage of the release reports that the manifest carries a sorry_count of zero, meaning the formalisations leave no declared proof holes, with dependencies limited to standard logical axioms.

This changes what kind of claim the announcement is. When OpenAI disproved the Erdős unit-distance conjecture in May, validation ran through nine external mathematicians who read the argument and signed off, according to Tech Times. That is a real standard, and a social one. It scales with the supply of qualified readers willing to spend a week on someone else's press release. A Lean certificate scales differently. Anyone with a laptop can rebuild the project and watch the kernel accept or reject each line, checked against mathlib, the community-maintained library of roughly 210,000 formalised theorems assembled over nine years by mathematicians at universities worldwide. OpenAI produced the proofs. The thing that checks them was built by the field.

The enthusiasm came fast. Writing on X, the widely-followed AI account Chubby called the release evidence that "scientific reasoning is becoming a genuine model capability," arriving faster than most expected. The cost figure carried the argument further: OpenAI puts the token spend for the successful runs at roughly $2,000 at Sol API rates. Priced that way, an attempt on a decade-old open problem becomes a line item. OpenAI researcher Sébastien Bubeck has since framed the ten as illustrative rather than exhaustive of what the model has produced, as reported by AI Weekly.

Where the doubt sits

Nobody credible is arguing the builds fail. The objection is narrower and harder to resolve.

A Lean certificate proves that the argument establishes the theorem as stated inside Lean. Whether that formal statement captures the problem as the field understood it is a separate question, and formalisation involves choices about definitions, quantifiers and hypotheses. A subtly weakened statement can compile cleanly while the historical conjecture stands untouched. Checking the correspondence is human work requiring specialists across ten subfields, and no single reader can credibly assess the whole set. Lean also has nothing to say about novelty. It cannot tell you whether a result is already implicit in prior work.

The second objection concerns the denominator. Simon Willison, writing about the release, noted the missing complement: "No news on how many problems they spent $2,000 on without reaching a solution." The published figure covers runs that worked. It leaves out failed attempts, parallel search, the compute consumed during exploration, and the cost of training Astra. The reasoning walkthroughs share the shape. They narrate dead ends inside successful runs, so the released material documents the search behaviour of winners.

Terence Tao supplied the framing for this in a post on 15 December 2025, before the release. What current systems display, he argued, is artificial general cleverness rather than intelligence: "the ability to solve broad classes of complex problems via somewhat ad hoc means," stochastic or brute-force, fallible, sometimes uninterpretable, and valuable when coupled with stringent verification that filters out the wrong answers. Every clause of that describes the announcement.

The institutional objection is larger. The Leiden Declaration, published on 2 June 2026 and endorsed by the International Mathematical Union, drafted by 16 researchers across 15 universities and signed by Tao, Peter Scholze, Kevin Buzzard and Scott Aaronson among more than 3,000 others, names five risks from AI in mathematics: unreliable results, missing citations, dependence on closed commercial systems, exaggerated claims, and loss of scientific independence. The Lean certificates answer the first and fourth well. They leave the third where it was. Astra has no public release date, no API, no pricing. On OpenAI's own account, the most consequential mathematics of the year came out of a system nobody outside the company can run.

OpenAI addresses part of this directly. The company states that attribution should reflect how a result was produced, says humans prepared the manuscripts and formalised the proofs in Lean, takes responsibility for correctness, and declines to claim human authorship of the arguments. Chapter acknowledgements credit named mathematicians as critical readers: Popa on Connes rigidity, and Henry Bradford, Michael Chapman, Alon Dogon and Francesco Fournier-Facio on non-sofic groups. Consultation, on the company's own framing, rather than independent certification.

The read for anyone building on this

Three claims can hold simultaneously, and most of the argument online consists of asserting one as if it settled the others. The proofs are probably correct.

The transferable part concerns problem shape. A reasoning model earns its inference cost where the statement is precise, the solution space has structure, and a verifier is cheap relative to the cost of generating candidates. Lean is that verifier here, which is why $2,000 bought ten results instead of ten confident manuscripts. Search plus a checker that cannot be argued with is a strong pattern, and this release is the best evidence for it so far.

That boundary has not moved. Where no cheap verifier exists, where telling a good answer from a plausible one needs a domain expert, the same search produces expensive failures at scale. What moved is the ceiling on the correct side of the line, which now sits higher than the sceptical prior assumed.

Four things would settle the open questions, and none requires trusting anyone. An independent audit of whether the ten Lean statements match the informal theorems, which needs no cooperation from OpenAI and is the fastest signal available. Specialist assessment of novelty in each subfield. A published list of the problems Astra failed. And access to the model, so the result can be reproduced by someone other than its owner.