AI Solves 5 of 10 Research-Level Math Proofs

OpenAI has released its attempt at the First Proof challenge, a math test that requires AI to generate checkable, end-to-end proofs for complex, domain-specific problems. Unlike standard benchmarks, these problems involve sustained reasoning and expert-level argumentation. The model was tested on all 10 problems, with internal estimates suggesting at least five attempts (problems 4, 5, 6, 9, and 10) are likely correct. One initial claim (problem 2) was later withdrawn after community review. The results highlight how frontier math challenges can push AI beyond cookie-cutter benchmarks, testing its ability to handle ambiguity and produce scrutiny-proof work. OpenAI is training a new model focused on rigorous, extended thinking, and this exercise served as a proving ground. Notably, the model improved over just a few days, solving harder problems as it trained. Why it matters: This signals that AI is moving toward doing actual research-level work, not just solving textbook problems—a key step for automated scientific discovery.