Mathematics serves as a critical benchmark for assessing AI development due to its objective, verifiable nature. Epoch AI launched FrontierMath in November 2024, a comprehensive benchmark with progressive problem tiers, initially featuring 300 complex math questions. This resource aims to measure AI’s mathematical reasoning capabilities. Earlier models solved under 2% of these problems, but top models like GPT-5.2 now tackle over 40% of tier 1-3 questions and over 30% of tier 4. Recent breakthroughs, such as Google DeepMind’s Aletheia achieving PhD-level results autonomously, underscore the urgent need for updated benchmarks. In response, the First Proof challenge, launched by distinguished mathematicians, tests advanced AI in solving original, challenging problems. Although results varied, with no AI fully solving all 10 posed questions, efforts like this, along with FrontierMath: Open Problems, signify a shift in evaluating AI in mathematics, striving to align AI challenges with human mathematicians’ interests.
