OpenAI revealed 722 manuscripts of mathematical work from an unreleased frontier model, covering 372 result families, and AGMAI stated that the model had solved “hundreds” of open problems.
The work builds on the company’s earlier claim that the model had solved more than 100 longstanding mathematical problems, but the response was not limited to praise. Mathematicians raised concerns about ethics, the model’s training data, and the seemingly rushed way the work was announced, prompting questions about whether OpenAI values winning more than the norms of the field.
When Mathematical Progress Becomes a Credibility Crisis
OpenAI revealed work from an unreleased model through 722 manuscripts, covering 372 result families, and stated that it had solved “hundreds” of open mathematical problems, including one Millennium Prize Problem. Work at this level should have been good news for the field, but questions were instead raised about how it was announced and communicated to mathematicians.
The company therefore established the independent AGMAI expert panel to help oversee the presentation and publication of the results. Researchers, however, still question how much authority the panel has, whether it can represent the field, and whether OpenAI will genuinely listen. This lack of clarity has caused scientific progress to be viewed through the lens of a credibility crisis rather than as something to celebrate.
Where Is OpenAI Taking Mathematics?
OpenAI is positioning itself as an AI company using an unreleased frontier model to tackle longstanding mathematical problems. The company published 722 manuscripts covering 372 result families and claimed to have solved “hundreds” of open questions.
This direction is not unique to OpenAI, as Anthropic and other AI labs have also announced advances in mathematical problems. But OpenAI has moved forward so prominently that questions have followed—about how the work was announced, ethics, and the impact on a field that prioritizes creating knowledge over competing to win.
From Announcing More Than 100 Problems to 722 Manuscripts
In September, OpenAI stated that its model had solved more than 100 longstanding open problems spanning nearly every branch of mathematics. The latest collection reports work in 722 manuscripts, divided into 372 related result families.
| Factor | September announcement | Latest collection |
|---|---|---|
| Number of problems | More than 100 | Hundreds |
| Scope of work | Nearly every branch of mathematics | Covering 372 result families |
| Number of manuscripts | Not specified | 722 |
| Grouping of results | Not specified | 372 related result families |
The article does not disclose the name of either the previous or newer model, stating only that the latest collection came from an unreleased frontier model.
The numbers do not mean that the model solved the problems conclusively, because much of the work still needs to be verified by mathematicians—especially the 722 manuscripts organized into 372 result families, which represent a major burden for the expert community.
Another issue is the average reasoning time, equivalent to three hours of ChatGPT Pro. This figure shows how much compute is required to generate the answers and raises further questions about cost-effectiveness and practical use.
OpenAI also provided summaries of the reasoning process, compute estimates, and experimental statistics. But these details will have meaning only if others can transparently verify and understand the origins of the results.
How OpenAI Differs from the Mathematical Community’s Approach
The article reflects that OpenAI views mathematics as a competitive arena, while mathematicians prioritize advancing the field. Other AI companies are mentioned only as also announcing mathematical breakthroughs.
| Factor | OpenAI | Mathematicians | Other AI companies |
|---|---|---|---|
| Goal reflected in the article | Focused on winning | Focused on advancing the field | Announcing breakthroughs |
| Relationship with the field | Seen as a well-resourced newcomer | Grounded in existing research and norms | Mentioned as breakthrough announcers |
Success Accompanied by Questions About Transparency
OpenAI claims that its unreleased model produced work across 722 manuscripts, covering 372 result families, and solved one Millennium Prize Problem, while beginning to disclose more about its reasoning and compute usage.
But repeated flawed announcements have caused the achievement to be questioned, including the process, training data, unpublished work, and the actual authority of the independent committee of mathematicians.
Pros
- +Covers a large number of open mathematical problems
- +Solves one Millennium Prize Problem
Cons
- −Repeated flawed announcements and an unclear process
- −Ongoing questions about training data, unpublished work, and the committee’s influence
The Price the Mathematical Community Must Pay, Even Without a Product for Sale
The first cost is the burden of verifying a large volume of work. OpenAI disclosed 722 manuscripts and claimed to have solved “hundreds” of longstanding open questions. Mathematicians must therefore spend time determining which results are correct and genuinely verifiable.
Another form of damage is to trust. Rushed and repeatedly flawed announcements have raised questions about publication norms, as well as how much influence the independent committee actually has.
There are also concerns about training data and unpublished work. Andreas Thom alleged that conversations with ChatGPT may have contributed to results in his field, while OpenAI acknowledged that the work relied heavily on Thom’s and Gábor Kun’s earlier research.
Can the New Committee Really Repair the Relationship?
This independent advisory group consists of leading mathematicians and is tasked with advising OpenAI and other AI companies on working with the mathematical community, including how to present and publish new results.
Researchers see it as a good starting point, but it remains unclear how much authority the group has and whether OpenAI will actually follow its recommendations. Another question is whether a small group of members can reflect the perspectives of the broader mathematical community. The relationship will improve only if the recommendations influence decision-making, rather than being used merely to create the appearance that the company is listening.
The Non-Sofic Groups Problem and Allegations About Training Data
Andreas Thom expressed concern that conversations between him, his colleagues, and ChatGPT before the announcement may have contributed to OpenAI’s success on the non-sofic groups problem. He accused the company of being opaque and acting “dishonestly” regarding the origins of its training data.
OpenAI acknowledged that the result relied heavily on earlier work by Thom and Gábor Kun, but the article has not proven that those conversations were used to train the model or directly contributed to the result. The clear facts at this point are therefore limited to the allegations, concerns, and the company’s acknowledgment.
What to Watch After This Wave of Work Is Published
OpenAI stated that it has 722 manuscripts covering 372 result families and has solved more than 100 longstanding open problems. This disclosure could accelerate discovery, but it also increases the burden on mathematicians to verify a large number of results while addressing questions of ethics and transparency.
The independent committee of mathematicians is a first step, but it still needs to prove how much influence it will truly have and whether OpenAI will listen. Ultimately, will this body of work advance the field, or add to the burden of verification and the problems of trust? The promise to learn from mistakes must be demonstrated through real action.