An internal version of OpenAI's next-generation Astra models has demonstrated an impressive result: 10 new solutions and significant progress on long-standing open problems in mathematics and theoretical computer science. There is no public access to the model yet, but the published materials are already sparking active discussion in the scientific community.

Details of the breakthrough

The company has released a 249-page technical report, detailed breakdowns of the model's reasoning process, and a repository with formalized proofs in Lean 4—a system for machine verification of the correctness of logical steps. According to my analysis, the key feature is that the mathematical arguments were fully generated by Astra, with humans only assisting in preparing the manuscripts. At the same time, OpenAI honestly states: attributing authorship of a proof to a human would distort the system's actual contribution.

The cost of computational resources for finding solutions is estimated at approximately $2000 based on Sol API rates. However, this figure does not account for the costs of verification and preparation of results, nor the undisclosed number of failed attempts.

Quantum complexity and post-quantum cryptography

Results related to quantum topics deserve special attention. Astra proved exponential parallel repetition for arbitrary finite quantum games—this could be significant for quantum verification and cryptographic protocols. A stronger result was also obtained regarding the complexity of the closest vector problem, which underlies lattice-based cryptography—one of the main directions in post-quantum protection.

It is important to emphasize: these achievements do not imply an immediate threat to already deployed post-quantum systems. This concerns fundamental theoretical results that could influence future developments.

Other achievements

Among other claimed breakthroughs are the construction of a non-sofic group, improved estimates for sphere packing in high dimensions, new lower bounds in arithmetic circuit complexity, a counterexample to the Conn rigidity conjecture, and results on the Ehrhart conjecture. This is fundamental mathematics whose practical impact will not be immediate.

Verification and scientific integrity

Formalization in Lean reduces the risk of logical errors but does not replace scientific expertise. Mathematicians will need to assess how accurately the formal statements correspond to the original problems. OpenAI references the Leiden Declaration on AI and Mathematics, emphasizing transparency of attribution, but the date of Astra's public launch and the possibility of independent testing remain undisclosed.

This breakthrough continues a series of AI achievements in mathematics: previously, an OpenAI model disproved the 80-year-old Erdős conjecture, and an Anthropic researcher found a counterexample to the Jacobian conjecture using Claude.

My view: This publication is an important signal for the entire industry. If the results are confirmed, we will witness AI transitioning from a tool for researchers to an independent generator of new scientific knowledge. For the crypto industry, this means the need to reconsider approaches to mathematical security, especially in the context of post-quantum cryptography.