OpenAI drops 700+ AI math papers

Top mathematicians can spend an entire career trying to crack just one or two foundational problems. OpenAI just dropped hundreds at once.

OpenAI published 722 mathematical manuscripts to a GitHub repository, grouped into 372 result families that the company claims solve or advance major open problems. The work was produced by an unreleased internal frontier model across fields like number theory, geometry, and theoretical computer science.

The real kicker is the compute efficiency. OpenAI posed roughly 4,000 problems to the system. While its September Navier–Stokes attempt reportedly mobilized 10,000 agents for days, the company claims nearly all of these new results required just one prompt, averaging the compute equivalent of three hours of ChatGPT Pro thinking.

Lean proofs and prime numbers

The boldest claim in the dump is a proof advancing the quasi-Riemann hypothesis, a weaker version of the famous million-dollar problem tracking how prime numbers are distributed across numbers.

While the model did not crack the full Riemann hypothesis, it showed that non-trivial zeros cannot exist in the outer 12.5% of the critical strip, shrinking the potential territory by 25%.

To keep researchers from dismissing the papers as elaborate hallucinations, OpenAI leaned on Lean, a formal verification language that lets computers verify proofs line by line. The release catalogue includes 162 papers whose primary theorems are fully formalized in Lean, giving scientists mathematically verified proofs they can run on a machine right now.

Another manuscript tackled Hilbert's Tenth Problem, asserting that even when allowing rational numbers as solutions, there is no universal algorithm to decide whether an arbitrary polynomial equation has solutions, drawing theoretical boundaries around computation itself.

Outrage, awe, and a $5M grant

The math community’s reaction was immediate and deeply fractured.

Levent Alpöge, an Anthropic mathematician behind a prominent Jacobian conjecture advance in July, called the drop "obviously the most significant moment in mathematical history."

Yet hours before, WIRED reported widespread anger among academic mathematicians. Many claimed OpenAI had previously promised to space out its releases to avoid drowning peer review in machine noise. An OpenAI spokesperson responded that the company was unaware of any such agreement.

To mend fences, OpenAI posted an update on October 15 pledging $5 million to a Simons Foundation initiative. The grant will fund independent conferences and workshops to help researchers evaluate the dump, along with dedicated model access for participating mathematicians.

Why it matters

We think this release signals a structural shift: mathematical discovery is no longer bounded purely by human insight, but by chip hours.

For researchers and computer scientists, the immediate value is proof of concept. Having 162 machine-checked theorems means this is not speculative marketing; it is functional, verifiable mathematics produced at scale. But the barrier to entry is shifting. Human mathematicians will spend less time cranking through brute-force steps and more time curating problems, steering search parameters, and interpreting what the model actually found.

OpenAI played a smart hand by coupling unreleased model output with Lean verification. By providing computer-checkable files, they bypassed typical AI skepticism.

However, the dump strategy is reckless. Leaving roughly three-quarters of the papers unformalized forces an unpaid peer-review burden onto human academics who never asked for it. If future releases continue at this scale, scientific publishing will either need automated Lean pipelines at the door, or the sheer volume of AI output will break the academic verification system entirely.