OpenAI's Math Repository: What Verification Looks Like in Practice

AI-generated editorial illustration; not a screenshot.

The Snapshot

On October 6, 2026, OpenAI shared AI progress in mathematics with a public GitHub repository, produced by an unreleased internal model. The model is not named, and I do not substitute a product name for it.

The figures below are a dated snapshot: the repository README and history at commit fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb, used for the original post and rechecked on October 10, 2026. Repositories change, so check the current state before you quote any number.

• 719 manuscripts across 372 result families. Families group related manuscripts. This is not 719 independently solved open problems.
• Approximately 4,000 research problems were posed to the model after existing math evaluations saturated. No success rate follows from that figure.
• Topics include π, graph coloring and the complexity of matrix multiplication.
• OpenAI reports about 42% of top-line results formalized in Lean (300 of 719 in the repository history). Lean lets a computer check a formal proof.

The launch README listed 722 manuscripts. Three were withdrawn on October 7, which is why this pinned snapshot counts 719.

The Correction Log Is the Interesting Part

On October 7, OpenAI withdrew three manuscripts after a sign error and the dependent arguments that rested on it. It also revised 14 others. The repository history records mixed corrections: proofs, statements, hypotheses, dependencies and citations.

You can read a few things from that. A public record of withdrawals means someone is checking. It also means the catalogue changed within a day of release after three manuscripts were withdrawn, which is what honest correction looks like and also why the number should be dated. And the repository ships supporting artifacts alongside papers: proof code, Lean formalizations, ten selected reasoning summaries and compute estimates. The summaries are selected, not all raw reasoning, and the compute figures are not elapsed time or a bill.

What These Numbers Do Not Establish

Be careful with three inferences:

1. Formalized does not mean independently reviewed. A Lean result shows that a stated formal claim has a checkable proof. A comparator on a supporting result does not establish that a paper's main theorem is right. OpenAI's percentage is its own statement of coverage; I have not rerun the audit.
2. A catalog is not a verdict. I have not solved, checked or independently confirmed any of these results, and I have not run the mathematics.
3. OpenAI also says it will fund workshops and programs to help the mathematics community understand AI-produced results, and that it is working toward responsible model release. These are stated plans, with no dates or amounts to quote.

I want to see how mathematicians check, challenge and build on this work. That is where results become foundations.

A Verification Pattern You Can Borrow

I run several agents on parallel projects. The practical lesson I take is about handoffs. Every handoff from an agent should arrive with something checkable:

• A test that passes or fails.
• A trace that shows what was tried.
• A source record for every factual claim.
• A clearly stated result, with its scope and its known gaps.

Then add the part this repository demonstrates well: a public change log. When a result is withdrawn or revised, the record shows why. In a business setting, that log is your audit trail.

For selecting the right checks, see building an evaluation harness before you ship. For the question of what to log in production, see AI observability: what to log.

How to Ask a Vendor About Verification

Use this as a short list in any AI procurement conversation:

• What artifact accompanies each output (test, trace, source)?
• What share of results has automated verification, and what does that verification not cover?
• How are errors discovered, recorded and announced?
• Who outside the vendor can inspect the evidence?

Read Next

The original post is on LinkedIn. Primary sources: OpenAI's announcement, the pinned README, the catalogue, the history and the launch README. The Agent companion, Verifiable handoffs between agents, covers the operations side. The week's wider context is in the October 2–8 recap, and a related question about engineering for trust is in the Frontier Academy article.

If you want verification designed into an agent workflow from the start, see our AI agent development service. You can also write via the contact page.