OpenAI’s unreleased Astra system has reportedly generated solutions to 10 longstanding problems in mathematics and theoretical computer science, with each argument formalized as a machine-checkable Lean 4 proof. Tech Times reported that the work was released on Aug. 1 with technical manuscripts and proof files published under an Apache 2.0 license, although Astra itself remains unavailable to the public.

The reported results cover group theory, operator algebras, combinatorics, high-dimensional geometry, quantum complexity, circuit complexity and lattice-based cryptography. The most prominent is a construction claimed to provide the first known example of a non-sofic group, answering a question that has remained open since mathematician Mikhail Gromov introduced sofic groups in 1999.

Sofic groups can, in a precise mathematical sense, be approximated by finite permutation systems. Many major classes of groups are known to have this property, but mathematicians had not established whether every countable discrete group was sofic. The Astra construction would settle that question negatively if the formal statement accurately represents the established problem and the broader mathematical community accepts the construction.

Another claimed counterexample addresses a rigidity conjecture associated with Alain Connes. That problem asks whether certain von Neumann algebras preserve enough information to identify the groups from which they were constructed. Other reported contributions include three results related to problems posed by Paul Erdős, an improved general upper bound for high-dimensional sphere packing, a theorem concerning repeated two-player quantum games, circuit-complexity lower bounds for the permanent and work relevant to lattice cryptography.

The use of Lean 4 is central to the announcement. Lean is an open-source proof assistant that represents mathematical statements and arguments in a formal language. Its relatively small trusted kernel checks whether every step follows from the declared definitions, axioms and previously established results. A completed proof can therefore be checked by running the relevant files rather than relying exclusively on manual review.

That verification has an important limitation. A successful Lean build establishes that the encoded theorem follows within the formal system; it does not, by itself, establish that the theorem entered into Lean perfectly matches the informal open problem recognized by specialists. Mathematicians must still inspect the definitions, assumptions and correspondence between the formal and conventional statements. They will also need to evaluate the conceptual value and originality of the constructions.

Tech Times reported that the proof repository is publicly available and can be checked independently using Lean 4. Public certificates would make the core validity claims more reproducible than AI-generated arguments presented only in ordinary mathematical prose. They would also allow researchers to identify precisely which assumptions a result uses and where supporting lemmas enter the proof.

Thomas Bloom, a University of Manchester mathematician who maintains an online catalogue of Erdős problems, described the results on X as “big news,” the report said. OpenAI mathematics research head Sébastien Bubeck also publicly confirmed the work, while researcher Noam Brown noted that Astra had not solved any Millennium Prize Problems.

The reported release follows growing use of formal proof systems as a benchmark for mathematical AI. Lean’s community-developed mathlib supplies a large library of definitions and proven theorems on which automated systems can build. Formalization remains labor-intensive, but it provides a common verification layer that does not require readers to trust a model developer’s assessment of its own output.

Astra is described as a multi-agent model family designed for long-running tasks, with a coordinating agent delegating parts of a problem to subordinate agents before combining their work. OpenAI has not announced a public release date or clarified how Astra will fit into its existing product lineup. Until specialists examine the theorem statements and reproduce the proof checks, the certificates offer a path for verification rather than a substitute for mathematical review.

Sources: Google DeepMind