OpenAI published a new collection of mathematical results produced by an internal frontier AI model, offering a closer look at how advanced AI systems may contribute to mathematical research and scientific discovery.
The company released the results publicly through a GitHub repository, allowing mathematicians and researchers to examine the problems, proposed solutions, and supporting materials.
OpenAI said it consulted with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study while developing its approach to releasing the work. The consultation was intended to help establish a responsible framework for sharing mathematical results generated with advanced AI systems.
A significant portion of the work has also been formalized using Lean, a programming language and theorem prover that allows mathematical arguments to be checked computationally.
These formalizations provide an additional layer of verification by translating mathematical proofs into a form that software can systematically validate.
OpenAI said many of the proofs in the initial collection already have Lean formalizations and that additional formalizations will be added as they become available.
Alongside the mathematical results themselves, OpenAI released information intended to provide greater transparency into how the internal model produced the work.
The supporting materials include 10 summaries describing aspects of the model’s reasoning process, estimates of the computational resources used, and statistics showing how many mathematical problems were attempted during the research process.
OpenAI said the average result required an amount of computation roughly equivalent to three hours of ChatGPT Pro thinking.
The release is part of the company’s broader effort to understand how frontier AI models can contribute to areas where progress requires extended reasoning, experimentation, verification, and specialized domain knowledge.
Mathematics is particularly useful for evaluating these capabilities because proposed results can often be checked rigorously through traditional peer review, computational verification, or formal proof systems such as Lean.
OpenAI also plans to support the broader research community as mathematicians evaluate AI-generated discoveries.
The company said it intends to fund workshops, conferences, and special programs focused on understanding significant mathematical results produced with the assistance of artificial intelligence.
These programs could provide researchers with opportunities to independently evaluate results, explore new methods for human-AI collaboration, and develop standards for assessing AI-generated mathematical work.
OpenAI is also working toward responsibly releasing the internal frontier model that generated the results.
The company plans to continue testing frontier AI systems across mathematics and other scientific disciplines as it develops models and tools intended to assist researchers with increasingly complex scientific problems.
The broader objective is to determine where advanced AI can meaningfully contribute to scientific discovery while developing verification methods that allow researchers to assess the validity and significance of AI-generated findings.
