OpenAI published 722 mathematical manuscripts covering 372 groups of results, claiming they contain solutions to “hundreds” of open problems from an unreleased frontier model.
This body of work still awaits detailed review by mathematicians and assessment of its impact. At the same time, it raises questions about ethics, how results are announced, and how human researchers should be credited. OpenAI published the materials on GitHub, along with solution approaches and research references.
What is this body of work, and how should it be read?
OpenAI released the work through GitHub, comprising 722 manuscripts organized into 372 related groups of results, along with solution approaches and research references.
It should be read as work that the mathematical community is still reviewing and assessing—not as a final conclusion—because mathematicians need time to understand these results.
When answers to mathematical problems arrive faster than the review process
This news is exciting because OpenAI published 722 manuscripts covering 372 groups of results and stated that they include solutions to “hundreds” of open problems. However, these figures do not yet mean that every answer is correct.
Mathematicians still need time to verify, understand, and assess the impact of the entire body of work. The speed of the announcement may therefore be outpacing the review process, which is why this news should be read cautiously.
Where OpenAI places this body of work in the company’s mathematical research journey
This publication follows an announcement in September that an OpenAI model had solved more than 100 previously open mathematical problems. What was initially just a claim has now been accompanied by manuscripts and some additional details published through GitHub.
More broadly, this is part of OpenAI’s rapidly growing body of mathematical work, alongside research from rival laboratories such as Anthropic. The field is therefore still in the process of gathering information, verifying results, and understanding how much these works may change research methods.
From the original announcement to 722 manuscripts
In September, OpenAI said that its model had solved more than 100 long-standing open mathematical problems across multiple fields. This publication provides the details in the form of 722 manuscripts and 372 groups of related results.
| Factor | September announcement | This publication |
|---|---|---|
| Problems mentioned | More than 100 | “Hundreds” |
| Scope of information | Covering multiple areas of mathematics | 722 manuscripts, 372 groups of results, summaries of reasoning, compute estimates, and experimental statistics |
| Publication format | Company announcement | GitHub, with solution approaches and references |
This publication gives the community a more concrete basis for beginning to examine the work, but the source has not concluded that all of the results have been verified.
What the numbers and details reveal about the process
This news is not only about the answers; it also reveals the process: there are 722 manuscripts organized into 372 groups of results, with claims covering “hundreds” of open problems. When reading it, it is important to consider both the scope of the work and how it is grouped, rather than simply counting the manuscripts.
OpenAI also provided summaries of the model’s reasoning, estimates of computational resources, and the number of problems attempted. The “average result” reportedly took an amount of thinking equivalent to three hours of ChatGPT Pro reasoning. These details help provide a better basis for evaluating the process, but the work still requires further review by the mathematical community.
OpenAI and Anthropic: What the source confirms and what cannot yet be compared
Both OpenAI and Anthropic have produced more mathematical work, while the field is still reviewing and processing these results.
| Factor | OpenAI | Anthropic |
|---|---|---|
| Mathematical work | More work has been produced | More work has been produced |
| Status in the field | Still being processed | Still being processed |
The source does not provide enough detail about Anthropic, and there is no shared standard for measuring the quality or quantity of the work. It is therefore not yet possible to determine who has produced superior results.
Strengths of this publication and questions that remain unanswered
Publishing the manuscripts through GitHub, together with solution approaches and references, allows people in the field to follow up and conduct further verification. OpenAI also provided process-related information, such as reasoning summaries, compute usage, and the number of problems attempted.
However, the model used has not yet been disclosed, and the impact of the work still awaits assessment by mathematicians. There are also questions about ethics, how the work was announced, and how the mathematicians whose work formed the foundation for these results should be credited.
Pros
- +Manuscripts published through GitHub
- +Solution approaches and references provided
- +Process-related information disclosed
Cons
- −Model name has not been disclosed
- −Impact cannot yet be assessed
- −Raises ethical and mathematician-credit concerns
The costs that do not appear in the computational figures
The important cost is not just compute, but also the time mathematicians must spend reviewing and understanding all 722 manuscripts, which cover 372 groups of results. The impact of this work therefore still awaits assessment by the mathematical community.
AGMAI warned that using mathematical work as a marketing tool could damage the field. In addition, the speed of the announcement increases pressure on the company to disclose the model name, prompts, and computational costs through appropriate academic channels.
What to watch after the publication
The next step is to see how mathematicians review and understand these results, particularly their correctness and impact on the field.
OpenAI said it would improve the manuscripts through citations, mathematical explanations, and presentation, while providing clearer disclosures of the model name, prompts, and computational costs. The debate over academic practices, ethics, and credit for the mathematicians whose work lies behind these results is also unlikely to end anytime soon.