OpenAI's unreleased Astra model cracks three more Erdos problems
An unreleased OpenAI model produced a counterexample to Erdos's 1946 unit-distance conjecture in May 2026 — the first historically significant AI proof — and on August 1 an unreleased model named Astra added 10 more advances, including three Erdos problems. Thomas Bloom's erdosproblems.com, a hobby list of nearly 1,000 problems he built with ChatGPT, is now the de facto benchmark both DeepMind and OpenAI push against. The site's comment section flattened the field: Terence Tao and non-professionals like Wouter van Doorn trade proofs on the same page, and problems 333 and 728 fell to early-career mathematicians Kevin Barreto and Liam Price using GPT-5.2 Pro. The flip side is that some professional mathematicians are stepping back from these problems as less-specialised contributors start landing real results.
Source: quantamagazine.org ↗
These models are "changing dramatically the way mathematical research is being done," said Noga Alon of Princeton University, who has solved dozens of Erdős problems over his decades-long career.
Why this matters
- → AI models now solve historically significant math problems, shifting research from human specialists to automa
- → Democratized problem-solving: non-professionals and early-career mathematicians compete with Fields medalists
- → Professional mathematicians stepping back from classical problems as AI and amateurs land real results.