OpenAI publishes 722 math manuscripts from an unreleased internal model, claiming progress on top open problems
2026-10-06
OpenAI published 722 math manuscripts, grouped into 372 families of related results, from an unreleased internal frontier model, in a public GitHub repository alongside a blog post, saying it consulted the Institute for Advanced Study's independent Advisory Group on Mathematics and AI before release. The collection came from evaluating about 4,000 research problems and used an average of roughly three hours of ChatGPT Pro compute per result; the model itself was not released. Commentators point to claimed progress on a quasi-Riemann hypothesis result, a faster-than-n-log-n integer multiplication result, a uniqueness result for the elastic inverse problem open since 1994, and a claimed resolution of Barnette's Conjecture (problem 180). Sam Altman called it 'a new era of discovery.' An independent analysis estimated about 20% of the results are disproofs or counterexamples. OpenAI's Will Depue said he expects some results won't survive scrutiny and built a site to track which human papers the release cites; François Chollet asked whether the gains generalize beyond easily-verifiable math/code domains.
Significance 4: A high-profile capability claim spanning hundreds of results including claimed progress on decades-old open problems — lab-reported, self-selected, and not yet independently verified, so capped at 4 per the rubric rather than 5.
Implications · machine-drafted, not owner judgment
Whether or not every claimed result holds up, OpenAI timed and framed this release as evidence for its AGI narrative — exactly the kind of high-profile capability claim the tracker's AGI-declaration-credibility thesis is built to test. Math results that are easy to verify once stated are also the domain most favorable to RLVR-style training, so independent verification of even a handful of headline claims would be a stronger signal than the sheer volume of manuscripts.
- Independent mathematician verification of the quasi-Riemann, Barnette's Conjecture, or integer-multiplication results within a month or two
- A second lab publishing a comparable verified-math result or disputing OpenAI's claims directly
- Any retraction or correction within the 722-manuscript set
Sources
- curator [AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers retrieved 2026-10-07
- curator Quoting Jake Boggan retrieved 2026-10-07