AI Solves Open Mathematics Problem for First Time on FrontierMath Benchmark
Key Takeaways
- ▸An AI system solved a previously unsolved mathematical conjecture from a 2019 paper, marking the first solution on Anthropic's FrontierMath: Open Problems benchmark
- ▸Multiple leading AI models (GPT-5.4 Pro, Gemini 3.1 Pro, and Opus 4.6) demonstrated capability to solve the problem, suggesting progress across the AI industry
- ▸The solution is slated for peer-reviewed publication with potential for novel mathematical insights, demonstrating AI's emerging role in advancing frontier research
Summary
Anthropic's FrontierMath: Open Problems benchmark has achieved a historic milestone with an AI system successfully solving one of its curated research problems—a mathematical conjecture that human mathematicians had been unable to solve since 2019. The problem, contributed by mathematician Will Brian, was originally from a paper co-authored with Paul Larson and had resisted multiple solution attempts over several years. Kevin Barreto and Liam Price first elicited the solution using GPT-5.4 Pro, with Geby Jaff independently solving it shortly thereafter.
The breakthrough demonstrates that current frontier AI models can tackle genuinely difficult, unsolved mathematical problems that lie at the edge of human mathematical research. Multiple state-of-the-art models—including Gemini 3.1 Pro, GPT-5.4, and Opus 4.6—have been confirmed capable of solving the problem through Anthropic's testing scaffold. Will Brian plans to publish the solution, potentially including novel follow-on work inspired by the AI's approach, with the initial solvers offered coauthorship on any resulting papers. This achievement represents a significant validation of FrontierMath as a meaningful benchmark for measuring AI progress on problems of genuine mathematical significance.
- FrontierMath: Open Problems serves as a credible benchmark for measuring AI progress on genuinely difficult, real-world mathematical challenges
Editorial Opinion
This achievement represents a meaningful inflection point in AI's capability to contribute to frontier mathematics. While solving a single moderately-interesting problem does not revolutionize mathematics, the fact that multiple leading models can tackle previously unsolved research problems—with solutions deemed publishable by human mathematicians—suggests AI is transitioning from a tool for optimization and pattern-matching to a genuine research collaborator. The key question now is whether this capability scales to harder problems and other domains, and whether such AI-assisted breakthroughs become common or remain rare exhibitions of frontier capability.
