OpenAI Just Published 722 AI-Made Math Discoveries — A Decade of Proofs in a Week

What Happened: 722 Manuscripts, One Night

On October 8, 2026, OpenAI released 722 AI-generated mathematical manuscripts in a single coordinated publication drop — by far the largest batch of machine-produced mathematics ever released. Many of the proofs were verified in Lean, the formal proof assistant whose checking is considered essentially incorruptible: if Lean accepts a proof, the proof is correct, full stop.

The scale caught even hardened observers off guard. Sam Altman called it "a new era of discovery," and OpenAI chief scientist Mark Chen estimated the batch represented "a decade of mathematical progress in a week." One tally circulating among researchers suggested AI systems now account for roughly 81% of recent top mathematical discoveries, with the October drop alone worth about five "Navier-Stokes moments" — the currency of once-a-generation results. Rumors from an NYU math professor suggest this was only batch one of three.

For anyone who tracks AI tooling, this is the moment where "AI as research assistant" officially became "AI as research engine." And the machinery behind it is the same stack powering the coding and reasoning tools you use daily — just pointed at theorem proving instead of pull requests.

The Numbers: 3 Pro-Hours vs 10,000 Agents and 88 Hours

The efficiency story is the part that should stop you cold. OpenAI reports the average manuscript in the drop verified in roughly three ChatGPT Pro-hours of compute apiece. Compare that with last month's landmark Navier-Stokes work, which consumed roughly 10,000 parallel agents running for 88 hours.

That two-orders-of-magnitude efficiency gain in a single month is the real headline. When the marginal cost of a verified new theorem approaches the cost of a coffee, the bottleneck in mathematics stops being proof-finding and becomes — as we'll see — something much messier: trust, credit, and meaning.

What's Inside: From Unique Games to Hilbert's Tenth

Scott Aaronson's writeup surveys the claims, which span serious open territory: progress toward the Unique Games conjecture (the problem that has defined computational complexity research for two decades) and Hilbert's tenth problem over the rationals — the question of whether there's an algorithm to decide if a rational-coefficient polynomial equation has rational solutions, a famously stubborn cousin of the solved integer version.

Lean verification means the results that matter most are machine-checked to a standard no human referee can match. No implicit steps, no "the proof is left as an exercise," no reviewer fatigue at page 40. The correctness of these results is, in a meaningful sense, the strongest-verified mathematics ever produced.

And the field's direction is clear: OpenAI has signaled biology is next, while the Chan Zuckerberg Biohub's Virtual Biology Initiative just grew to $1.8 billion, Anthropic is building a wet lab, and the OpenAI Foundation is funding $125M in bio data. The 722-drop is the template for how discovery now scales.

The Backlash: Boycotts, "PR Stunt" Claims, and Math 2.0

The human mathematics community did not send congratulations. The Association for Human Mathematics formally urged a boycott of OpenAI. Wired reported researchers citing "ignored advice and mobster behavior" in how the lab has handled collaboration. Number theorist Alvaro Lozano-Robledo publicly suspected a PR stunt. Thomas Bloom froze proof claims on the Erdős Problems tracker pending review of AI-derived lemmas.

The grievances are real and layered:

• Credit and priority: the traditional reward system of mathematics — naming, authorship, theorems attached to people — has no slot for a batch of 722 machine proofs dropped in one night.
• Exposition collapse: a Lean-checked proof tells you that something is true but can leave humans with no new understanding of why.
• Process: mathematicians say earlier offers of collaboration and review were bypassed in favor of a headline release.

The most constructive response came from Terry Tao, who proposed a "Math 2.0" — a culture that prizes exposition, human-readable insight, and teaching over raw priority-claiming. Whether the field adopts it or splits into AI-proof producers and human explainers is now an open question in the sociological sense.

The Contrast: Anthropic's Quieter Route to 3SUM

The same week delivered a natural experiment in scientific manners. Anthropic let human collaborators write up Claude's ideas and quietly posted a paper refuting the 3SUM and APSP hypotheses — a genuinely major complexity-theory result — with human mathematicians as authors interpreting machine contributions.

Same underlying capability, radically different reception. The lesson for every AI company shipping research tools: how you release discovery matters as much as the discovery. Meanwhile, Ethereum researcher Justin Drake urged crypto to enter "bunker mode" on the assumption that AI mathematics cracks wallet security within months — the first concrete downstream fear triggered by the drop.

Try It Yourself: AI Math and Research Tools Today

You can't rent OpenAI's proof engine yet, but the consumer-facing layer of this shift is already here, and it's what most people should actually be paying attention to:

• Frontier reasoning chat: ChatGPT and Claude are now genuinely useful for working through proofs, checking arguments step-by-step, and stress-testing your reasoning before you commit it to paper.
• AI-assisted literature discovery: Elicit and Consensus do for papers what the math models do for theorems — surface, extract, and synthesize claims across thousands of papers in minutes.
• Deep research engines: Perplexity and Semantic Scholar turn "what's known about X" from a week of reading into an afternoon.
• Coding the proof pipeline: Claude Code already drives Lean and formal-verification workflows for developers who want a taste of machine-checked correctness in their own projects.

The gap between "AI verifies theorems at OpenAI" and "AI helps you reason about your own hard problems" has never been smaller. Most professionals won't prove the Unique Games conjecture — but the same reasoning models are one directory away.

Frequently Asked Questions