Epoch AI has introduced a new secret benchmark called FrontierMath, designed to evaluate the advanced mathematical reasoning capabilities of AI models. This benchmark has proven to be extremely challenging, with top AI models like GPT-4o achieving success rates of less than 2%. The benchmark also poses significant difficulties for PhD-level mathematicians. FrontierMath focuses on testing complex mathematical reasoning and understanding, rather than the ability to rephrase answers from sources like Stack Overflow. The results highlight the limitations of current AI systems in solving novel and creative mathematical problems, challenging their real-world capabilities.