OpenAI has released results and data it says were generated by a new AI model, resolving many long-standing open mathematics problems.
On October 6, 2026, The Verge AI reported that OpenAI had disclosed a body of mathematics produced by an unreleased frontier model: 722 manuscripts, grouped into 372 result families, each family collecting papers that share a conclusion or a proof route. OpenAI did not name the model, disclose parameter counts, training data composition or per-problem inference cost, and did not say what proportion of the 722 manuscripts had been checked by a proof assistant such as Lean. The disclosure format is itself notable. OpenAI skipped the usual rhythm of one paper per announcement and released the entire batch at once, closer in shape to a results inventory than to a single claim
Three lines explain the timing. The first is model-side reasoning. Since 2024, frontier models have become markedly more stable over long chains of inference, shifting from producing answers to producing inspectable proof steps, which is the precondition for mathematics to be generated at batch scale. The second is the maturity of formal tooling. Lean and its Mathlib library have accumulated enough theorem coverage for machine-generated proofs to be automatically checked, the hardest available test of whether such results hold. OpenAI did not disclose th
Liu Gong holds that the question worth asking about this batch is not whether the results are right, but who gets to confirm that they are.
From a technical standpoint, the bar in machine mathematics has moved from can it solve to can it be checked. When AlphaProof solved four of six problems at IMO 2024 in July 2024, the decisive step was translating natural-language solutions into Lean code for machine verification. That translation, not the model's fluency, is what made the result acceptable. OpenAI has now released 722 manuscripts and 372 result families without publishing formal verification coverage, leaving the hardest link in the dark. Volume is not evidence of quality. If most of the 372 families are corollaries of a handful of larger results, the academic increment may be far smaller than the headline number suggests. If they are genuinely independent, their worth depends entirely on whether each family survives scrutiny. Liu Gong expects that the figure which ultimately determines the historical weight of this release is a third-party replication ratio in Lean, not the total of 722.
On the business side, mathematics generates almost no direct revenue, yet it is the most expensive billboard a frontier model can rent. Mathematics has no noise, no subjective scoring, and machine-checkable right and wrong answers. It is one of the very few domains where the claim that a model is genuinely smart can be converted into a verifiable proposition. That credibility transfers, and it transfers into the paid territory: code generation, research assistance, drug and materials design, all of which depend on long chains of reliable inference. The cost structure, however, is deeply asymmetric. OpenAI's marginal cost for generating these manuscripts is largely a compute bill. The cost to mathematics of digesting them is a large pool of scarce expert hours distributed across subfields. That asymmetry will persist, and it will sit at the center of any future negotiation between universities and labs.
For the field as a whole, the pressure lands on institutions. The three rules mathematics has run on for a century, authorship, priority and peer review, were designed for human authors producing at human speed. When a company releases 722 manuscripts at once, no review system can physically absorb them, and priority cannot be established the traditional way, because the results have no citable publication and no author declaration that passed through an editorial process. Liu Gong predicts two responses within a year. Mathematical journals and preprint platforms will write disclosure and labeling rules for machine-generated output. Separately, Lean verification coverage will be treated as a front-end filter standing in for the first stage of peer review, with the community screening for checkability before deciding which results deserve human reading. Neither route is cheap, but both are cheaper than leaving 722 manuscripts in the limbo of the not-yet-disproved.
- 722 manuscripts map to 372 result families, about 1.9 manuscripts per family, indicating a structure built around problem families rather than individual papers.
- Against AlphaProof's four of six problems solved at IMO 2024 in July 2024, this release is more than two orders of magnitude larger by result-family count, though the two are not comparable benchmarks and cannot be converted into a capability multiple.
- An active mathematician publishes roughly 2 to 5 papers a year, so 722 manuscripts are on the order of the annual output of 150 to 350 math
The impact splits in two. Within three months, the immediate change is procedural: journals, preprint platforms and university mathematics departments will be forced to answer whether submissions must declare model involvement, whether machine-checkable verification is a precondition for review, and how priority is registered for machine output. Within a year, if third-party replication rates come in high, formal verification shifts from tooling to infrastructure, and Lean coverage plus verification compute become public-resource questions. If replicatio
Reactions inside mathematics follow the split seen in earlier rounds. One camp reads the 722 manuscripts as an unambiguous positive signal: proofs that fail machine checking will eventually be filtered out, and whatever survives is net gain. Another camp focuses less on correctness than on cost and order, asking OpenAI to publish the model identifier, the mapping between manuscripts and result families, and formal verification coverage, without which the community cannot judge how many people to assign to reading. Community discussion centers on authorsh
Questions worth exploring:
- How machine-produced mathematics should be attributed and how priority should be registered, and whether existing academic norms need rewriting
- Whether Lean verification coverage can substitute for the first stage of peer review as a screening gate in high-volume fields
- How the bargaining relationship between universities and frontier labs shifts when generation costs fall far below review costs
This item comes from The Verge AI (published 2026-10-06 23:26:38). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.