OpenAI says an internal version of its upcoming AI model, known as Astra, has achieved ten new research breakthroughs in mathematics and theoretical computer science, tackling problems that had remained unsolved for at least a decade.
The company announced the results in a research publication, accompanied by a 249-page technical manuscript and machine-verifiable Lean 4 proof certificates for each breakthrough. OpenAI also released the formal proof files on GitHub, allowing researchers to independently verify the mathematical results.
According to OpenAI, the advances span multiple branches of mathematics and theoretical computer science, marking one of the company’s most significant research announcements to date.
Astra Tackles Long-Standing Mathematical Problems
Among the reported achievements is what OpenAI describes as the first explicit construction of a non-sofic group, a long-standing open question in group theory since mathematician Mikhail Gromov introduced the concept of sofic groups in 1999.
The company also reported progress on several other notable problems, including Connes’ rigidity conjecture on von Neumann algebras, Ehrhart’s volume conjecture, and three problems from mathematician Paul Erdős’ well-known collection of unsolved questions.
OpenAI further said Astra produced new results in high-dimensional sphere packing, quantum game theory, and arithmetic circuit complexity, demonstrating capabilities across several advanced research disciplines.
OpenAI Publishes Verifiable Proofs
OpenAI’s Head of Mathematics Research, Sebastien Bubeck, described the results as “beautiful” in a post on X, noting that each breakthrough includes both a machine-checkable Lean certificate and a detailed reasoning walkthrough.
The company estimated that generating all ten research results required computing resources equivalent to roughly $2,000 using Sol API pricing, highlighting what it says is the model’s efficiency on complex mathematical research tasks.
The publication of Lean certificates addresses one of the key concerns surrounding AI-generated mathematics: independent verification. Because the proofs can be checked using the Lean theorem prover, researchers can validate the results without relying solely on OpenAI’s claims.
Research Community Watches Closely
The announcement comes amid growing debate over artificial intelligence’s role in mathematical research.
Earlier this year, a group of mathematicians released the Leiden Declaration, expressing concerns about AI-generated research, the use of published mathematical work for model training, and the increasing number of research claims announced through company publications instead of traditional peer-reviewed journals.
OpenAI has previously reported research advances using the same long-horizon model family. In May, the company said an earlier version of the model produced a proof related to the Erdős unit distance problem, drawing praise from several researchers, including Fields Medal-winning mathematician Tim Gowers.
Mathematician Thomas Bloom, who maintains the erdosproblems website, described the latest announcement as “big news” and suggested the new results could prove even more significant than OpenAI’s earlier mathematical breakthrough.
OpenAI has not announced when Astra will become publicly available, describing it only as its next major AI model. While some industry observers have speculated that Astra could become part of a future GPT series, the company has not confirmed those reports.
Whether the broader mathematics community ultimately accepts the findings will likely depend on independent verification and future peer review. For now, OpenAI’s publication represents another step in the growing role of AI systems in frontier scientific research.
