While the global community debates the ethical frameworks of artificial intelligence, we are witnessing a historic moment in fundamental science. On August 1, OpenAI presented proofs for ten mathematical problems that had remained unsolved since 2016. This is not just another language model update—it is a demonstration of how AI is beginning to work at the forefront of human knowledge.

What lies behind the breakthrough

The solutions were obtained using an internal version of the Astra model—the next generation of OpenAI's architecture, positioned as a separate class alongside Sol, Terra, and Luna. In my assessment, the key innovation here is not in the computations themselves, but in the methodology: the model not only generated hypotheses but also formalized each proof in the Lean language for machine verification. This is a critically important step, as it eliminates the main problem of AI research—hallucinations and logical errors.

The economics of the process are also interesting: according to OpenAI's calculations, the total cost of tokens spent on finding the solutions would have been about $2000 at Sol API rates. This is negligible compared to the years of work by mathematical teams, opening new horizons for the democratization of science.

Breakthrough results

Among the key achievements is a construction proving the existence of non-sofic groups. This closes the central question that had stood since 1999, when Mikhail Gromov introduced the concept of soficity. Also disproved were Connes's rigidity conjecture on von Neumann algebras, the Ehrhart hypothesis on volume, and Erdős's problem No. 183 on multicolor Ramsey numbers. I would particularly note the improved estimates for sphere packing in high dimensions up to the Cohn–Elkies threshold—this has direct applied significance for coding theory and cryptography.

Mathematician Thomas Bloom from the University of Manchester has already called these results "big news" and a "significant step" in the field of constructions. I agree with him: this is comparable in significance to the recent disproof of a 1939 conjecture by Anthropic's Claude model, which we observed in July.

However, it is worth maintaining a sober perspective. Astra could not approach the "Millennium Problems"—the seven problems for which the Clay Mathematics Institute promises $1 million each. This is a reminder that even the most advanced AIs are still far from solving the deepest mathematical mysteries. OpenAI has already submitted the model for evaluation to regulators in Washington—under the new rules of the Trump administration, before it becomes publicly available, possibly as GPT-6 or within the GPT-5 lineup.

My verdict: we stand on the threshold of a new era where AI becomes not a tool, but a full-fledged partner in scientific discoveries. The fact that the model handles tasks that people have struggled with for decades is a signal for the entire industry: investments in reasoning models will pay off not only in commercial applications, but also in fundamental science.