OpenAI says a new internal model that began training on August 28 has done far more than crack the Navier–Stokes problem: the company claims the system has now resolved more than 100 long-standing open problems across most areas of mathematics.
In a new post announcing an independent mathematics advisory group, OpenAI says the pace of the model’s progress “has surprised the mathematicians within OpenAI” and has triggered internal debate over how to prepare the broader mathematical community for capabilities that are arriving faster than expected.
The company is careful to frame the claims as significant but still requiring care. Mathematics, OpenAI argues, is not just another benchmark domain; new proofs can feed into applications across science and engineering, which makes responsible development and deployment “important beyond mathematics itself.”
The announcement lands after OpenAI said an unreleased internal model—described in earlier coverage as significantly more capable than GPT-6 Astra—had produced a machine-checked solution connected to the Navier–Stokes existence and smoothness problem, one of the Clay Mathematics Institute’s seven Millennium Prize Problems. That episode immediately became a test case for the field: not only over whether the proof would survive independent scrutiny, but over how results should be credited, verified and communicated when the line between human and machine discovery becomes blurred.
OpenAI acknowledges the tension directly. The post references a recent open letter, “A Severe Misalignment of AI in Mathematics,” in which mathematicians warned about the negative externalities of using solved open problems as a scoreboard for AI systems. The concern is not merely academic. If frontier labs begin treating centuries-old conjectures as marketing milestones, the risk is that verification, attribution and the slower work of human understanding get trampled by competitive pressure.
That is the backdrop for the new advisory group. OpenAI says the body will act as a bridge to the mathematical community and the wider public, giving mathematicians a formal channel to weigh in on emerging results before they are released into a hype cycle.
The group’s mandate is notable for its breadth. It will advise OpenAI on the review and communication of new results, help assess their significance, coordinate dissemination, and comment on academic and professional standards for mathematical research. It will also advise on how OpenAI’s tools can support mathematical research and learning—an implicit nod to the fact that the company wants these systems used by working mathematicians, not only flaunted as frontier benchmarks.
OpenAI is also trying to pre-empt the obvious criticism that the group will be a fig leaf. The company says the advisory group will operate independently, will be free to offer unsolicited advice, can comment publicly on OpenAI’s impact on mathematics, and can make its advice public. Members will not be paid by OpenAI, and the group can change membership as it sees fit. Most strikingly, OpenAI says the group will not advise the company on how to pace internal mathematical progress—an attempt to separate governance of disclosure from the cold logic of capability development.
The initial members are heavyweight: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten and Melanie Matchett Wood. The group is hosted at the Institute for Advanced Study, a symbolic choice given IAS’s place at the center of pure mathematical research.
The recent history of AI mathematics has already shown how brittle trust can be. Earlier episodes around GPT-5 and Astra included claims that were later corrected, clarified or criticized for overstating what had actually been achieved; Terence Tao has warned that AI-generated proofs can become a net negative for mathematics if they arrive as opaque black boxes rather than legible contributions. Against that backdrop, OpenAI’s promise of independent advice looks less like philanthropy and more like risk management.
There are also unresolved questions around the Navier–Stokes result itself, including how it relates to independent work by mathematicians using AI tools and how much priority should be assigned when models, agents and humans all contribute to a proof. OpenAI says the advisory group will not settle every dispute. But it may set a template: if AI labs want to claim breakthroughs in fields built on verification, they may need to accept a new layer of external scrutiny over how those breakthroughs are announced.
For now, OpenAI’s message is measured but unmistakable: the company believes its models are entering a phase where they can generate novel mathematics at scale. The next test is whether the mathematics community—and the public—can be brought along.