OpenAI Publishes 722 AI-Generated Math Papers Claiming Riemann Breakthrough (2026)

On October 6, 2026, OpenAI announced on X that it has published 722 mathematics manuscripts produced by an unreleased internal frontier model. The OpenAI math manuscripts are organized into 372 result families and are available in the public GitHub repository openai/math under the Apache 2.0 license, many accompanied by formal proofs written in the Lean proof assistant. It is the largest public release of AI-generated mathematical research to date, and several of its headline claims would have made waves even as the output of a large human collaboration.

According to OpenAI’s post “Sharing AI progress in mathematics,” the project started with roughly 4,000 mathematical problems posed to the internal model. Each finished result represents about three hours of reasoning at the level of ChatGPT Pro. The company framed the release as a step toward transparency around advanced AI capabilities.

That framing has drawn scrutiny. The Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, issued recommendations on September 29, 2026 asking laboratories to publish the prompts behind such results and to refrain from using them as marketing. Scientific American reported that OpenAI shared compute statistics about the project but declined to release the prompts themselves.

What the papers claim

The most striking claim is a zero-free half-plane Re(s) > 7/8 for Dirichlet L-functions, described as a “quasi-Riemann hypothesis.” To be clear, this is not a proof of the full Riemann hypothesis, one of the most famous unsolved problems in mathematics. The full hypothesis requires all nontrivial zeros of the Riemann zeta function to lie on the critical line Re(s) = 1/2. A zero-free region reaching 7/8 would nevertheless count as a major advance in analytic number theory, a field where progress is typically measured in small increments.

A second claim concerns matrix multiplication: the exponent for fast matrix multiplication algorithms is lowered from 2 to 9/4, or 2.25. Matrix multiplication sits at the heart of scientific computing and machine learning itself, so a better exponent would have wide theoretical interest.

A third result is a near n log n algorithm for integer multiplication, approaching the theoretical optimum for multiplying very large numbers. Algorithms of this kind matter for cryptography and computational mathematics.

One important caveat: the model that produced these results has not been released, so outside researchers cannot probe or reproduce the system that generated the manuscripts.

Why it matters, and what to be careful about

For readers, the significance is straightforward. AI is moving from helping mathematicians check their work to producing candidate research at scale. That shift is part of a broader set of AI trends reshaping science, and it means machine-generated mathematics may soon become a routine input to the field rather than a curiosity.

The verification story is mixed. Lean formalizations cover the main results of 162 of the papers, which gives those claims a machine-checked backbone. The remaining 560 papers have not been formally checked, and OpenAI itself warns that some could contain issues. On the encouraging side, an audit of the results posted to arXiv on August 1 found no confirmed substantive errors. Mathematicians are taking the release seriously: Rutgers professor Alex Kontorovich wrote on X that the achievement would merit “an instant Fields Medal, no questions asked” had a human done it. Coverage from TechSpot, Unite.AI, and Let’s Data Science has focused on both the scale of the release and the open questions around how the results were produced.

Whether all 722 manuscripts survive human scrutiny will take months to determine. For now, the release marks the moment AI mathematics went from promise to corpus, with the proofs, the code, and the caveats all in public view. For readers following the field, the papers themselves are now the story: they are public, checkable, and numerous enough to keep mathematicians busy for a long time.

Sources: TechSpot, Unite.AI, Let’s Data Science.