OpenAI publishes 722 math manuscripts and Lean proofs from an unreleased frontier model

On October 6, 2026, OpenAI published a research post presenting new results on open problems in mathematics produced by an unreleased internal frontier model. The company also shared Lean proof formalizations and research details in a public GitHub repository, openai/math, released under an Apache-2.0 license. According to Unite.AI, the release contains 722 manuscripts organized into 372 problem families, alongside the Lean formalizations and abridged summaries of the model's reasoning. Making the formalizations public lets outside researchers inspect the derivations themselves instead of relying on OpenAI's own account.

Early on October 7, TestingCatalog reported that the vast majority of the results came from the same procedure using the same unreleased internal OpenAI model, with each result taking roughly three hours on average. Unite.AI and The Verge both carried the same figures: 722 manuscripts covering 372 result families that group related papers.

Reporting from IT之家 added that AGMAI, a newly formed independent advisory group of mathematicians created to communicate the results responsibly, described the release as containing answers to "hundreds" of unsolved problems. The materials also include partial summaries of the model's reasoning, estimates of compute consumption, and statistics on how many problems were attempted. OpenAI said the compute behind a typical result was roughly equivalent to three hours of continuous thinking by ChatGPT Pro. In September the company had said its model solved more than 100 long-standing open problems spanning most branches of mathematics. For weeks beforehand, observers had been expecting the release, and OpenAI had not said which problems were solved or when the results would appear.

The Verge noted that the batch extends a run of breakthroughs that have both impressed and unsettled parts of the mathematical community, while raising questions about research ethics and scholarly conduct. The reports do not state any conclusions on those questions.