In early August, I gained access to materials confirming that OpenAI has made a breakthrough in the field of automated theorem proving. An internal version of the Astra model, which ChatGPT's developer positions as its next flagship system, successfully tackled ten mathematical problems that had remained unsolved since 2016. This is not just another step forward—it is a demonstration that AI can work at the level of a professional research mathematician.

The essence of the achievement and its costs

Based on OpenAI data, my estimates suggest that the total volume of computational tokens spent on finding solutions is equivalent to approximately $2000 at API rates for the Sol model. This is negligible compared to the value of the results obtained. The manuscripts were prepared by humans working alongside the model, after which Astra formalized each proof in the Lean language, which is used for machine verification of theorems. The company has made all certificates and reasoning records publicly available, which is crucial for verification by the scientific community.

Key mathematical victories

Among the most significant results is a construction proving the existence of non-sofic groups. This closes a central question that mathematicians had been unable to solve since 1999, when Mikhail Gromov introduced the concept of soficity. Astra also disproved Connes's rigidity conjecture on von Neumann algebras, solved the Ehrhart hypothesis on volume, and resolved Erdős's problem No. 183 on multicolor Ramsey numbers. The model obtained new lower bounds on the complexity of computing the permanent using arithmetic circuits and proved a theorem on parallel repetition for two-player quantum games. Additionally, upper bounds on sphere packing density in high dimensions were improved up to the Cohn–Elkies threshold, and bounds for binary codes were strengthened.

Community reaction and limitations

Thomas Bloom, a mathematician at the University of Manchester, has already called the published results "big news" and a "significant step" in the field of constructions. However, one should not think that Astra is omnipotent. The model failed to approach the "Millennium Problems"—seven questions that the Clay Mathematics Institute named as the most important in mathematics in 2000, promising $1 million for solving each. Noam Brown, co-author of Astra's reasoning technology, confirmed that attempts were made, but they did not yield success.

My analysis: This event marks AI's transition from a role as a tool for computations to a full-fledged partner in fundamental science. However, it is important to understand: solving open problems is not a replacement for human intuition, but rather its enhancement. For now, AI is not capable of generating fundamentally new paradigms of thinking, but it can already systematically comb through the space of hypotheses, saving researchers years of labor.