Via mathful.com
Epoch AI's benchmark for testing frontier AI models against research-level mathematics is growing faster than the models can keep up, and the implications for crypto-adjacent AI infrastructure are worth watching.
Epoch AI’s FrontierMath benchmark now includes roughly 50 unsolved research-level mathematics problems in its Open Problems track, and AI systems have managed to crack exactly three of them. That’s a 6% success rate against questions that have stumped professional mathematicians for years.
The benchmark represents one of the most ambitious attempts to measure whether AI can do more than regurgitate textbook solutions and can actually contribute to the frontier of human knowledge in pure mathematics.
What FrontierMath actually tests #
The Open Problems component draws from active research areas including combinatorics, number theory, algebraic geometry, and topology. These aren’t exercises with known answers sitting in the back of a textbook. They’re genuine open questions that working mathematicians haven’t been able to resolve.
One of the three problems AI has solved involved a Ramsey-style hypergraph problem, a category of combinatorial mathematics that deals with the conditions under which order must appear within chaos.
Each submitted solution goes through custom verification checkers to ensure authenticity. The system demands rigorous, independently verifiable proofs.
The broader FrontierMath suite also includes Tiers 1 through 4, a separate dataset of approximately 300 problems with known solutions. The Tier 4 set alone contains 50 problems and was completed in June 2025. The Open Problems track launched in early 2026 as the next logical step: testing AI against questions where nobody knows the answer yet.
The growth curve is steep #
In late 2024, frontier AI models solved nearly zero FrontierMath problems across the benchmark. By mid-2026, that figure climbed to 30-40% over roughly 18 months.
The benchmark has been expanding to maintain its relevance. Two additional problems were added in February 2026, and curation remains ongoing. The most recent site update was logged on June 27, 2026.
OpenAI is the sole purchaser of access to FrontierMath’s solution verification process and previously funded the foundational Tiers 1-4 dataset. Schmidt Sciences, the philanthropic initiative founded by former Google CEO Eric Schmidt, provides independent support for the Open Problems track specifically. That funding structure means the hardest tier of the benchmark operates with some independence from the very companies whose models it evaluates.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our