SWE-bench Author Calls Millennium Problems the Best AI Benchmark, Predicts 70% of Open Math Questions Solved Within a Year

Ofir Press, a Princeton researcher and author of SWE-bench, argued in an October 29 tweet that the best current AI benchmark is the Millennium Problems, far ahead of any other benchmark. Extrapolating from models' current success rates on hard math problems, he said predicting that "about 70% of open math problems will be solved within a year" is not an exaggeration. The claim is notable coming from the creator of SWE-bench: even the author of a software engineering benchmark admits math benchmarks better measure frontier capability.

Confirmed

• Ofir Press (Princeton researcher, SWE-bench author) believes the Millennium Problems are the best current AI benchmark by a wide margin
• Extrapolating from current solve rates, he predicts about 70% of open math problems will be solved within a year
• He argues math is more "concrete, definable, and suited to reinforcement learning" than other fields including software engineering, so other domains won't replicate math's rapid progress
• He marvels that 15 years ago, he wouldn't have believed that a consumer-grade domain-level math robot would arrive before consumer-grade self-driving cars
• He believes the "jagged frontier" of model capabilities will persist, continuing to advance in ways unintuitive to humans
• He further asks whether current progress will spawn new open problems an order of magnitude harder and unsolvable short-term, or reach some "math capability peak" where nearly all formalizable math problems are quickly solved

Why it matters

• If "70% of open math problems solved within a year" comes true, mathematical research will undergo a paradigm shift, and the value and scarcity of remaining problems will be repriced
• The gap between progress in math and SWE offers a clue to where RL works: formalizable domains with cheap verification break through first
• A persistent "jagged frontier" means AI capability forecasts can't be extrapolated evenly across domains; policy and application planning need domain-by-domain assessment

2026-10-29 ~ 2026-10-30 · 9 related posts

• [source] Ofir Press: Millennium Problems is AI's best benchmark; 70% of open math could fall in a year — OfirPress · 2026-10-29
• Ofir Press wonders if AI math progress hits a wall or a 'peak math ability' — OfirPress · 2026-10-29
• Ofir Press: Math is far more RL-able than SWE, other fields won't move as fast — OfirPress · 2026-10-29
• Ofir Press thread: math is more concrete and RL-able than SWE — OfirPress · 2026-10-29
• Ofir Press: AI's jagged frontier advances unintuitively — math robots beat self-driving cars — OfirPress · 2026-10-29
• Ofir Press: AI's jagged capability frontier will keep advancing in non-intuitive ways — OfirPress · 2026-10-29
• [source] Millennium Problems called AI's best benchmark; 70% of open math solved within a year? — DimitrisPapail · 2026-10-29
• [source] Ofir Press Predicts ~70% of Currently Open Math Problems Solved Within a Year — OfirPress · 2026-10-30

1 near-duplicate retellings: OfirPress