On January 18, 2026, Greg Brockman, co - founder of OpenAI, and researcher Neel Somani made an announcement on the X platform. They revealed that GPT - 5.2 Pro had successfully cracked the 281st unsolved number theory problem proposed by the renowned mathematician Paul Erdős. Since its introduction in 1980, this problem has consistently baffled the academic world.

GPT - 5.2 Pro employed an ingenious proof logic. It introduced tools from ergodic theory and combined Haar measure calculations with the Birkhoff ergodic theorem. In English academic expressions, when presenting a new approach or method, using “employed” instead of “constructed” here can better convey the idea of utilizing existing concepts in a novel way.

After thorough verification, Terence Tao, a winner of the Fields Medal, lauded it as the “clearest example of AI solving open mathematical problems.” This achievement marks a significant leap for AI, transitioning it from the realm of auxiliary computation to that of auxiliary proof. In Western academic culture, the Fields Medal is highly regarded, and having a Fields Medalist praise a result adds immense credibility.

At present, this solution has been categorized as a “Part 1” result obtained through a new method. The related verification system is structured as a three - tier mechanism. First, GPT - 5.2 Pro generates an initial draft. Then, formalization tools are used to translate the code. Finally, external experts conduct a review. This tiered approach is in line with common academic verification practices in the English - speaking world, ensuring the accuracy and reliability of the results.