OpenAI Releases 722 AI Math Manuscripts Across 372 Major Problems
OpenAI has released 722 manuscripts with claimed solutions and progress reports spanning 372 major math problems. Engadget
The manuscripts were posted in a GitHub repository that went live at 6 P.M. EDT on Oct. 6. OpenAI listed the release on its news page as a Research article titled "Sharing AI progress in mathematics" dated Oct. 6, 2026.
The results came from an unreleased ChatGPT pioneer model. OpenAI said the average result took about three hours of ChatGPT Pro use. Nearly every paper was produced from a single prompt given to a single AI agent.
The claims include a solution to the four-dimensional Kakeya conjecture in geometry and progress toward the Riemann hypothesis about prime numbers. The set covers claimed solutions and partial progress, not only complete proofs, across the 372 problems.
OpenAI did not release specific compute times for individual problems or the prompts it used. Without runtime, sampling budget, and prompt structure, independent reproduction from the release alone is limited.
This release follows two earlier mathematics disclosures listed by OpenAI. On Aug. 1, the company shared new results on long-standing open problems in mathematics and theoretical computer science, including geometry advances. On Sept. 8, it shared an AI-generated solution to the Navier-Stokes Millennium Prize Problem, with a writeup and a formal proof in Lean, software for computer-checked logic.
The broader context here is scale and checking. Hundreds at once move the hard work from writing to review. For mathematicians, the measure is not manuscript count but verification load per claim, and the difference between an argument people read and a derivation machines check.
The broader context here is also production method. Single-prompt, single-agent work with a few hours of Pro-level run time points to long automated runs rather than teamwork with a person. Researchers often tune helper setups and split tasks across agents. Without prompts and per-problem compute, that spare setup is hard to judge.
In my view, the Lean proof in the earlier Navier-Stokes release is more consequential than the manuscript count here. Informal manuscripts help only if referees can follow them. Formal proofs change the workflow, because computer checking absorbs part of the cost before human review. Worth flagging for labs and infrastructure teams: if automated mathematics continues, the scarce resource becomes trusted checking, formalization pipelines, and time for long jobs.
In my view of what comes next, faster work on hard conjectures is the payoff if even a fraction holds up. Failed attempts, partial lemmas, and progress reports have value when clear. The field now has a concrete collection to test that idea.
