OpenAI has published a batch of 722 artificial intelligence-generated manuscripts which, according to the company itself and the independent advisory group AGMAI, contain solutions to hundreds of open mathematical problems. The news is significant because it marks one of the largest volumes of mathematical results attributed to an AI system released to date, and because it comes amid a growing debate over how these advances should be communicated without becoming marketing tools.
What the batch of 722 manuscripts contains
The documents come from a frontier model that OpenAI has not yet released to the public. The material is organized into 372 families of results, a grouping that allows related articles to be gathered together, as a single finding may be developed across several different manuscripts.
Along with the texts, the company has shared summaries of the reasoning followed by the model, estimates of the compute used, and statistics on how many problems were attempted. According to OpenAI, the average result in the batch consumed compute equivalent to three hours of ChatGPT Pro reasoning.
Until now, the company had not specified which problems it had solved or when it would publish the details. In September, it had stated that its model had solved more than 100 long-standing open problems in most areas of mathematics. The new batch significantly expands on that initial figure.
AGMAI, the committee monitoring how these results are communicated
The Advisory Group on Mathematics and Artificial Intelligence, known as AGMAI, is an independent advisory group composed of top-tier mathematicians. Its role is to help labs like OpenAI communicate these findings responsibly.
At the end of September, AGMAI published its first recommendations. It asked companies to release results quickly and, whenever possible, through standard academic channels. It also demanded transparency regarding the name of the model used, the prompts employed, and the associated computing costs.
The committee included an explicit warning: companies must avoid treating the publication of mathematical results as a marketing tool to promote their models. AGMAI believes that such a practice can cause significant harm to the mathematical community.
GitHub as a publication channel and the doubts it raises
OpenAI has opted for a GitHub repository for this release, with its own protocols for review and citation of the articles. The company acknowledges that it continues to explore alternatives managed by the mathematical community itself that better align with AGMAI's guidelines.
This implies that the chosen channel is not a peer-reviewed journal, which is precisely the path the committee identified as preferred. OpenAI itself admits that this first batch is a starting point and commits to improving aspects such as citations, mathematical exposition, and the general presentation of results in future releases.
That commitment to improvement is revealing: it suggests that the published material does not yet reach the level of polish typical of a mathematical article intended to be read and cited by the scientific community.
An autumn full of announcements that the discipline is still digesting
This batch joins a series of recent announcements from OpenAI and rival labs, such as Anthropic. Among them are results linked to a Millennium Prize problem, one of the most famous open questions in mathematics.
The speed with which these companies have broken into the field this year, and especially the way they have presented their announcements, has opened an intense debate about research practices and ethics. One of the central questions is how to recognize the work of human mathematicians upon which these systems rely to produce their results.
Technology-focused media have noted that the real impact of this batch will take time to be felt, as mathematicians must still evaluate and verify the published material. A manuscript that took the model just three hours of compute may require weeks or even months of human review before being confirmed as valid.
How to interpret this announcement if you work with mathematics
It is important to distinguish between what OpenAI claims and what is actually proven. That a batch contains solutions to hundreds of problems is, for now, a statement from the company and its advisory committee, not a fact independently validated by the mathematical community.
Those who need to rely on any of these results should first check which of the 372 families it belongs to and which specific manuscript within that family contains the complete proof.
It is also recommended to consult the compute estimates and the reasoning summary that accompany each result, as AGMAI considers that information an essential part of responsible disclosure.
What it implies for researchers and the general public
For researchers, doctoral students, or mathematics professors, the OpenAI repository offers material subject to evaluation, but it does not constitute a source that can be cited without prior verification. The review of proofs remains a human task, and the burden of checking 722 documents now falls on the academic community.
For those who do not work in mathematics, the relevant data is different: a model that is not yet available to the public, with an average cost equivalent to three hours of ChatGPT Pro per result, is already generating texts that a committee of independent experts considers serious enough to create a specific oversight mechanism. When that model finally reaches the public, it will be possible to compare these claims with the actual use that each user can make of it.
Frequently Asked Questions
What is AGMAI?
AGMAI, an acronym for the Advisory Group on Mathematics and Artificial Intelligence, is an independent advisory group formed by top-tier mathematicians whose goal is to help AI labs communicate mathematical results generated by their models responsibly.
Have the 722 manuscripts already been verified by human mathematicians?
Not completely. The mathematical community still needs time to review and verify the material, and some experts warn that this process can take weeks or months for each manuscript.
Where have these manuscripts been published?
OpenAI chose a GitHub repository, with its own review and citation protocols, instead of an academic peer-reviewed journal, which is the channel AGMAI identified as preferred.
Which OpenAI model generated these results?
The manuscripts come from a frontier model that OpenAI has not yet released publicly, so users cannot yet verify its capabilities directly.
How much compute did it cost to generate these results?
According to OpenAI, the average result in the batch consumed compute equivalent to approximately three hours of ChatGPT Pro reasoning.
Why is there concern about how these mathematical advances are communicated?
Because AGMAI warns that some companies could treat these results as marketing vehicles for their models, a practice that the committee considers harmful to the mathematical community and to the recognition of the work of the human researchers involved.
