Cracking open conjectures becomes the new frontier benchmark, displacing competition-math scores
The measure of a frontier model is shifting from solving problems with known answers to resolving problems no one has solved. In one quarter AI has disproved the Jacobian and unit-distance conjectures, claimed a proof of the cycle double cover, and produced ten formalised results from Astra. Research mathematics is becoming the benchmark that separates the frontier from the pack.
The old benchmarks are saturating. When top models score near the ceiling on competition mathematics, the score stops discriminating, and the field reaches for a harder yardstick. Open conjectures are that yardstick: there is no answer key, no memorised solution, and success is unambiguous — the problem was open, and now it is not.
The strategic consequence is that frontier labs are racing on discovery, not chat quality. A model that can contribute to research mathematics is demonstrating exactly the long-horizon, verifiable reasoning that also underwrites high-value agentic and coding work. The math results are a proof of the underlying capability as much as an end in themselves.
The risk is over-reading a young signal. A handful of dramatic results, several still awaiting specialist review, is not the same as reliable research assistance across mathematics. But as a discriminator between the top models and the rest, 'can it move an open problem' has replaced 'can it ace a benchmark' — and that reframing is itself the story.
Quanta Magazine — The AI revolution in math has arrived → · Forbes — The AI breakthrough that has mathematicians paying attention →