Epoch AI’s FrontierMath benchmark has grown to 50 open research challenges in mathematics, with AI successfully solving three questions that have long baffled experts. A 6% breakthrough rate against unsolved problems signals a significant step forward in AI’s ability to push mathematical boundaries.
Benchmarking AI Against Unsolved Math
The FrontierMath Open Problems track tests AI models on genuine research questions from fields like combinatorics, number theory, algebraic geometry, and topology. Unlike textbook exercises, these problems lack known solutions, representing cutting-edge mathematics.
One of the solved challenges involved a complex Ramsey-style hypergraph problem, which explores how patterns inevitably appear amid chaos. Each AI proof undergoes rigorous verification through custom checkers to ensure authenticity and validity, mirroring standards expected in academic mathematics.
Rapid Progress and Growing Challenges
The benchmark’s history reveals an impressive trajectory. Near zero progress in late 2024 evolved into solving up to 40% of problems by mid-2026 in certain tiers. FrontierMath ranges from Tiers 1 to 4, with the initial 300 known-solution problems tackled before shifting focus to open questions in early 2026. Ongoing curation ensures the suite remains a true test of frontier AI capabilities, with updates continuing into June 2026.
OpenAI plays a critical role as the sole purchaser of FrontierMath’s verification pipeline, having supported earlier tiers as well. This work hints at broader impacts on crypto-related AI infrastructure and research tools, especially as AI starts tackling complex tasks beyond simple data regurgitation.
This material is for informational purposes and should not be considered financial advice.
