OpenAI's next GPT, the model behind a reported 10,000-agent Navier–Stokes proof, will ship soon

Sam Altman used a keynote in Abu Dhabi to preview OpenAI's next-generation GPT, telling the audience the model will ship "soon" and recounting the internal project that reportedly produced a proof of the Navier–Stokes existence and smoothness problem. According to a report on dev.to covering the appearance at the ADIPEC 2026 opening ceremony, 10,000 AI agents ran for 88 consecutive hours with no human in charge, then spent a further 17 hours converting the result into a machine-checked Lean formalization.

The dev.to post draws on Chinese-language tech coverage — Xinzhiyuan via NetEase and the Phoenix Tech AI Morning Post — and is, for now, the only account of the remarks. It also carries its own central caveat: none of the claimed results have yet been examined by outside experts.

How the swarm reportedly worked

The detail dev.to emphasizes is the absence of orchestration. There was supposedly no supervisory agent decomposing the problem into subtasks for workers to execute mechanically, the architecture most multi-agent systems use today. Instead, the agents selected their own directions, shared intermediate results, borrowed promising ideas from one another, and set their own next steps. The run is described as an autonomous push of nearly four days — a body of work the post suggests a human mathematician would struggle even to plan.

What the model reportedly produced

The headline claim is the first of the Clay Mathematics Institute's Millennium Prize problems to fall: Navier–Stokes existence and smoothness, the long-open question at the heart of fluid dynamics. The same account says the run's output reached well beyond a single theorem, touching the Riemann hypothesis and the Hodge conjecture, settling the Kakeya conjecture in all dimensions, overturning the exponential time hypothesis, and producing a proof of Yang–Mills existence. The total output is described as 722 research manuscripts rather than one paper.

The capability claim

Altman reportedly told the audience that "AI is now, most of the time, smarter than top scientists," and described the model behind the work as only a few months old. dev.to's take is that the line may read as hype, but the company has a track record of previewing real products from keynote stages, which makes the "soon" worth taking seriously as a shipping signal.

Still a single-source story

Every number above — 10,000 agents, 88 hours of proving, 17 hours of formalization, 722 manuscripts — comes from one secondary report relayed through Chinese-language outlets, and none of it has been independently corroborated. The strongest verifiable element, if it is ever released, would be the Lean formalization: a machine-checked proof can in principle be audited by anyone with the right tooling, which would settle questions of correctness even as debates over provenance and framing continue.

Why it matters

If the claim is even roughly accurate, it moves the frontier from models that assist mathematicians toward systems that organize and execute open-ended research on their own for days at a time. Discounted as showmanship, it still signals where OpenAI positions its next release: long-running, self-directed agent swarms aimed at problems that have previously defined entire mathematical careers. And because Altman says the model ships soon, the distance between the keynote narrative and a checkable artifact may close quickly. Until the Lean files or the manuscripts surface for review, the story stands as a milestone claim awaiting its audit.