OpenAI Posts 719 Math Manuscripts From an Unreleased Model and Withholds the Prompts

OpenAI published 719 mathematics manuscripts on Oct. 6 from an unreleased internal model, and kept back the prompts, the model's name and the compute spent on each problem. The advisory group it says it consulted had asked labs on Sept. 29 to disclose all three.

The collection sits in a public GitHub repository under the Apache 2.0 licence. Its README lists 719 manuscripts grouped into 372 families and says the collection "includes results at different stages of verification." None of the claims has been peer reviewed.

What was published and what was held back

OpenAI's blog post, which returned an access error to The Terminal, was summarised by Pulse 2 and LetsDataScience. LetsDataScience reports the model was given roughly 4,000 open problems and that the average kept result cost compute equivalent to three hours of ChatGPT Pro thinking. It also reports that an OpenAI spokesperson told Scientific American nearly all results came from a single prompt to a single agent.

The README says about 42 percent of top-line results have Lean formalisations, machine-checkable proofs, and that unformalised results "could have issues." Corrections will ship as new versions with earlier ones kept.

Item	Published	Withheld
Manuscripts	719 in 372 families	n/a
Reasoning summaries	10 of 372 families	The rest
Compute	Average of about three ChatGPT Pro hours	Per-problem figures
Prompts and model name	None	Both

The counts differ by outlet. LetsDataScience says 722 manuscripts and Gizmodo says 377 results, while the README says 719 manuscripts in 372 families.

What the advisory group asked for

The Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study has nine unpaid members, LetsDataScience reports, among them Fields Medalists Timothy Gowers and Martin Hairer and physicist Edward Witten. Its Sept. 29 recommendations, informed by more than 600 survey responses, asked AI companies to publish promptly, disclose the model, prompts and compute, and avoid using breakthroughs as marketing.

Gizmodo reports the group also said it does not endorse testing advanced problems on proprietary models the wider community cannot reach. The group said its role "should not be interpreted as a judgment of the impact of these results" and called the release "the beginning, not the completion," per LetsDataScience.

OpenAI spokesperson Lindsay McCallum Rémy told Gizmodo the group's advice has informed how OpenAI is sharing the results and that the company will keep taking community feedback. Gizmodo argues the new results come from the same unreleased model as a Navier-Stokes claim about a month earlier, so the group's request has not been followed.

The claims, and the first reactions

LetsDataScience lists a proof of the Unique Games Conjecture as entry 102 and a zero-free region above 7/8 for Dirichlet L-functions as entry 003. Gizmodo's reporter, Mike Pearl, says he is not qualified to judge the results. He sampled a paper claiming "the full Birch-Swinnerton-Dyer leading term formula," which Gizmodo notes holds only for elliptic curves over the rationals when a certain Selmer group has corank zero or one at some prime.

MIT mathematician Andrew Sutherland said the claims should be treated as unverified until the model can be replicated. "We should ask for receipts," he said, per LetsDataScience. Anthropic researcher Levent Alpöge called the release "the most significant moment in mathematical history" on X before adding caveats, the same report says. OpenAI chief executive Sam Altman wrote on X that he was looking at the stars "with extra awe."

OpenAI says it is working toward releasing the model responsibly and will fund workshops and conferences on AI-produced results. It gave no release date. The Terminal has also covered Rich Hickey's argument that AI is not a tool and a Berkeley study on AI use and persistence. Lean formalisations for the other 58 percent of top-line results are the next thing to watch.