OpenAI publishes report on AI progress in mathematics and opens the openai/math repository

OpenAI published a post titled "Sharing AI progress in mathematics" on its site and put the openai/math repository on GitHub. On October 13, 2026, the Two Minute Papers channel broke down the release in a video titled "OpenAI's New AI Is Doing Things We Barely Understand." The video description contains only two links to primary sources and no metrics. So below we separate what is known for certain from what remains open.

What was published

The available data indicates that OpenAI presented its mathematical results as a standalone publication and attached code to it. That detail matters. Companies usually show math performance through benchmarks in release posts, but here the topic gets both its own page and its own repository. The claim that the system does things "we barely understand" comes from the video's author. It is his wording in the title, not a statement from OpenAI.

The materials we have access to do not disclose:

• the name and architecture of the model that produced the results;
• the problem set and the share of problems solved;
• the compute budget and the running time per problem;
• how the solutions were checked: by a human, by a formal verifier, or both.

Without these parameters, any comparison with competitors is guesswork.

Context: where the interest in mathematics comes from

Mathematics has become the main proving ground for reasoning models. An answer is either correct or it is not, and a proof can be checked. In July 2025, an experimental OpenAI model scored 35 out of 42 points at the International Mathematical Olympiad, solving five of the six problems. That is gold-medal level. Since then, the question has shifted from olympiad problems to research problems: can a model find a new proof, or at least reliably verify someone else's?

A publication with code on GitHub is a logical continuation of this story. If the repository contains problems, solutions, and evaluation scripts, it lets you check the claims yourself. How complete the repository is remains unclear.

Why "barely understand" is a reason for caution

The phrase in the title can mean three different things. First: the model finds solution paths that are not obvious to humans. Second: interpretability is lagging, and nobody can explain why the reasoning works. Third: the title is clickbait. It is important for us not to merge these readings. Until we see conclusions from OpenAI itself, Two Minute Papers' assessment remains a commentator's opinion.

How to work with this

If you are a researcher or engineer, start with the primary sources, not the video. Check things in this order:

• open the post on openai.com and find the list of problems, the grading criteria, and the description of how solutions were checked;
• in openai/math, look at the license, the data format, and whether reproduction scripts are included;
• check whether there is formal verification, for example in Lean, or whether solutions were graded by hand;
• compare the stated compute budget with yours: a result achievable in hours on a large cluster is unlikely to be suitable for a product.

If you are building a product on LLMs, there is one practical takeaway. Math results are useful as an indicator of the ability to sustain long chains of reasoning, but they do not transfer directly to commercial tasks. Wait for independent verification before changing your stack or making promises to customers. We will return to this story when numbers from the post itself and the first reproductions appear.