OpenAI's Astra Model Solves Ten Long Standing Math Problems

OpenAI has announced that an internal version of its next major AI model, Astra, has produced solutions for 10 longstanding open problems in mathematics and theoretical computer science. The company made the announcement on August 1, publishing a 249-page manuscript and machine-checkable proof certificates for verification. The results, which span fields including group theory, cryptography, and combinatorics, represent a significant claim of AI's advancing capability in fundamental research.

The headline achievement is the first explicit construction of a non-sofic group, resolving a question posed by mathematician Mikhail Gromov in 1999. Other results include disproving Connes's rigidity conjecture in operator algebras, establishing new bounds on high-dimensional sphere packing, and resolving three problems from the late Paul Erdős's catalog. One finding directly concerns national security: Astra established that the closest vector problem, a lattice mathematics foundation for post-quantum encryption standards, is harder to solve than previously proven, potentially affecting cryptographic systems adopted by U.S. defense and intelligence networks.