The world of artificial intelligence has received yet another confirmation of its power: OpenAI's new flagship model, known under the codename Astra, has successfully tackled the proof of ten mathematical problems that had remained open since 2016. This is not just a technical triumph, but a signal that AI is transitioning from the role of a tool to that of a full-fledged researcher in fundamental science.

In my estimation, the key value of this achievement lies not in the theorems themselves, but in the methodology. The company did not simply force the model to produce answers; it built a pipeline: humans formulated draft reasoning, and Astra formalized each proof in the Lean language, which is used for machine verification of theorems. This means that every step of logic was verified with mathematical rigor, with no room for "approximation." All certificates and reasoning records are open to the community, making the process fully transparent.

Breakthrough results and the cost of the matter

Among the solved problems is a construction proving the existence of non-sofic groups, which closes a central question that had plagued mathematicians since 1999, when Mikhail Gromov introduced the concept of soficity. It also managed to disprove Connes's rigidity conjecture on von Neumann algebras, solve Ehrhart's conjecture on volume, and address Erdős's problem No. 183 on multicolor Ramsey numbers. Equally impressive is that the model obtained new lower bounds for the complexity of computing the permanent and proved a theorem on parallel repetition for quantum games.

The economic side of the matter is also noteworthy. According to OpenAI's calculations, the volume of tokens spent on finding answers would have cost approximately $2000 at API rates for Sol. This is a negligible sum compared to years of labor by human mathematicians, opening new horizons for the democratization of science.

Prospects and limitations

However, not everything is rosy. Astra failed to approach the "Millennium Problems"—seven problems for which the Clay Mathematics Institute offers $1 million for solving each. For now, this is beyond the capabilities of AI, suggesting that even the most advanced models will not replace human intuition on the deepest questions.

It is worth noting that on July 26, OpenAI CEO Sam Altman demonstrated Astra to politicians and regulators in Washington. This model may become the first to be tested under the new rules of the Trump administration, which require AI systems to undergo federal evaluation before public release. This hints that we are on the threshold of a new era of AI regulation, where scientific breakthroughs will go hand in hand with political decisions.

My verdict: Astra's success is a milestone that will redefine the relationship between AI and mathematics. But it is important to remember: the model is a powerful tool, not a replacement for the researcher. Its failure on the "Millennium Problems" is a reminder that true discoveries require not only computational power, but also creative courage, which algorithms do not yet possess.