OpenAI has published its proof attempts for First Proof, a research-level math challenge testing whether AI can produce correct, checkable proofs in specialized domains. The company ran an internal model on all 10 problems, which require building end-to-end arguments that experts must verify.

At least five proof attempts appear correct, including problems 4, 5, 6, 9, and 10, with several others under review. The company initially believed its attempt for problem 2 was likely correct but now acknowledges it is incorrect based on community analysis. The problems were authored by leading experts, with some remaining open for years before their solutions were found.

This work builds on OpenAI's earlier achievements, including gold medal-level performance on the International Mathematical Olympiad in July 2025 and a physics collaboration in November 2025 where GPT-5.2 proposed a candidate expression for a gluon-amplitude formula that was formally proved and verified. OpenAI researchers emphasize that novel frontier research represents the most important way to evaluate next-generation AI capabilities, as benchmarks can miss the hardest parts of research like sustaining long chains of reasoning and handling ambiguity.