OpenAI's mathematics meets reality: mass announcements and immediate retractions

On 6 October 2026, OpenAI announced a set of mathematical results produced by one of its internal frontier models, unreleased and unnamed. The manuscripts were deposited in the public GitHub repository “openai/math” under the Apache-2.0 licence. According to the documentation, most of the results came from submitting around four thousand problems to the model, with compute for the average result estimated at roughly three hours of ChatGPT Pro “thinking”. The initiative follows mathematicians' criticism of OpenAI's August claims, when the company asserted results on open problems, including work on Navier-Stokes. That criticism led to the creation of the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) at the Institute for Advanced Study, tasked with defining good publication practices.

OpenAI says it consulted AGMAI and partly followed its recommendations — publishing compute estimates and statistics, and announcing funding for future workshops — yet the release does not fully meet the group's criteria. The guidelines AGMAI published on 29 September 2026 call for results to be deposited in archives not controlled by the originating labs, and for models to be made public so the work can be reproduced. AGMAI also stated that it does not approve of submitting advanced mathematical problems to proprietary models (“we do not endorse this practice”) and reaffirmed that human understanding of mathematics remains paramount.

The fragility of such a massive release became apparent the day after launch. As of 7 October, the revision log records the retraction of three manuscripts: a sign error in a study on Weil classes invalidated the main argument and took two derivative papers down with it. The log also lists corrections to fourteen manuscripts and updated references in thirteen more. At present, 300 of the 719 main results, about 42%, are formalised in Lean. For the rest, OpenAI itself warns that some “may have issues”. As Annemarije de Boer noted in Yahoo Tech, formal verification alone is not enough to establish whether a result is genuinely new, significant or correctly framed. The results were selected by OpenAI, none appear to have undergone peer review, and the collection has not been independently assessed. Because the model is neither named nor accessible, the generation cannot be reproduced.

A release on this scale shifts a substantial share of the verification work onto the mathematical community: three retractions and dozens of corrections in a single day make that plain. Automating the writing is not the same as automating the truth. — Olya