On 20 October 2026, papers appeared in the latest arXiv listing in which mathematicians directly point to the help of models such as ChatGPT-6 Astra.
Among them is a proof of the Erdős–Szemerédi sunflower conjecture for any k ≥ 3. The authors note that the central theorem was found precisely by ChatGPT-6 Astra Pro.
Previously, only the case k = 3 was known, obtained in 2016 thanks to the breakthrough with cap-sets. Now an explicit improvement is claimed: the size of a family without a k-sunflower is bounded by a quantity less than 2^n.
In the same listing are other results involving AI: an estimate for the Borodin–Kostochka conjecture, an improvement in the running time for solving integer programming problems, and counterexamples to the Arveson conjecture in the commutative case.
Many authors write that the model proposed key steps or complete proofs, while humans verified and wrote them up. Some papers are accompanied by mentions of Lean checks, although the files have not yet been published.
All the claimed results await independent review. For now they remain in the status of “awaiting verification.”
