On September 8, 2026, OpenAI published a paper claiming that an internal, unreleased AI system had made progress on one of the seven Millennium Prize Problems: the Navier–Stokes existence and smoothness problem. The announcement got a lot of attention fast, partly because of the result itself, and partly because of how it was produced — a coordinated swarm of roughly 10,000 AI agents working in parallel.
As developers, we've all seen "AI solves impossible problem" headlines before, and most of them fall apart under scrutiny. This one is more interesting than most, but it also comes with real caveats that are worth understanding before you repeat the claim anywhere. Here's a breakdown of what was actually done, how the multi-agent workflow worked, and why the story isn't as clean as the headline suggests.
The problem
The Navier–Stokes equations describe how fluids move — they're used in weather forecasting, aircraft design, and blood flow modeling, among other things. They work great in practice. What's never been settled mathematically is whether, under the right conditions, a smooth three-dimensional fluid flow could spontaneously "blow up" — meaning some property like velocity shoots to infinity in a finite amount of time, breaking the model.
Proving whether that kind of breakdown, called a singularity, can or can't happen has been an open question since the 1930s. In 2000, the Clay Mathematics Institute put a $1 million prize on it, and it's remained one of the hardest open problems in math ever since.
OpenAI says its system produced a proof that a singularity can, in fact, form in finite time from a smooth, physically reasonable starting state — and that this proof has been formalized and checked in Lean, a proof-verification language that mathematicians use to catch logical errors that are easy to miss by eye.
How the multi-agent workflow worked
This is the part that's genuinely new, at least in scale. According to OpenAI's own writeup, the process looked roughly like this:
Starting in late August 2026, OpenAI trained a new internal model that outperformed its already-released flagship model on math benchmarks.
After hearing rumors that a rival lab might be close to solving a different Millennium problem, OpenAI decided to point this model at all of the unsolved Millennium Prize problems at once, plus a handful of other hard open problems.
Rather than running one long conversation with one model, they spun up large groups of coordinating agents. Each agent group could talk to other agents within its group, and different groups were given different framings of the same problem — some aimed at proving a statement true, others aimed at disproving it.
The group that eventually cracked the Navier–Stokes case grew to around 10,000 agents running at the same time.
Every so often, researchers used a coding-focused model to pull out the most promising partial results from different groups and feed them back in as new prompts, essentially cross-pollinating ideas between otherwise separate lines of reasoning.
Along the way, a smaller group of under 100 agents also worked out a related, slightly easier problem (a version of the same blow-up question for the Euler equations, which is what you get when you remove viscosity from Navier–Stokes). That intermediate result apparently helped point the larger effort in the right direction.
The full effort took about 88 hours of agent work, followed by another 17 hours to formalize and verify the proof in Lean.
Across the whole project, the agents exchanged close to 5 million messages and generated somewhere in the neighborhood of 300 billion output tokens.
So the "workflow" is closer to a massive, structured search process: many parallel attempts, deliberate diversity in approach, periodic synthesis of the best ideas, and a hard formal-verification step at the end to catch mistakes.
Why you shouldn't treat this as a settled result yet
A few things are worth flagging clearly:
It hasn't been independently certified. This is OpenAI's own internal evaluation, checked by their own Lean formalization. The Clay Mathematics Institute has not verified it, and OpenAI itself has said it isn't claiming the prize for this result. Outside mathematicians need time to review it properly, and that process is still ongoing.
There's an unresolved priority dispute. Around the same time, a mathematician at NYU and a researcher at a rival AI lab had been working for roughly a year on a closely related problem, using a mix of AI tools from both labs, and had reportedly made a breakthrough of their own just days before OpenAI's announcement.
OpenAI acknowledges it only started this specific push after hearing rumors about that other team's progress. The two results turned out to address related but not identical versions of the underlying question. Public statements from the NYU mathematician have raised pointed concerns about how OpenAI handled credit and communication during this period; OpenAI has disputed parts of that account. That dispute is still playing out publicly and isn't fully resolved.
Respected mathematicians are uneasy about the pattern. Some senior figures in the field have pointed out a broader worry: if AI labs increasingly treat famous open problems as marketing opportunities, and only publish final answers without the failed attempts and reasoning that usually teach the field something, it could quietly damage how mathematical progress actually happens — even when the individual results are correct.
What this means for practitioners
Setting aside the drama, the workflow pattern here is genuinely relevant to anyone building agentic systems:
Diversity over depth, at least initially. Running many independently-framed attempts in parallel found a promising lead that a single deep attempt might have missed.
Cross-pollination beats isolation. The step where useful partial results were pulled out of one group and fed into others was described as a turning point, not an afterthought.
Formal verification matters when correctness is non-negotiable. The Lean formalization step is doing real work here — it's the difference between "a model that sounds convincing" and "a proof that's been mechanically checked line by line."
Scale isn't free, and it isn't magic. Getting to this result reportedly took vastly more compute than OpenAI has used for past math results — and it still produced a result that overlaps with, rather than clearly surpasses, work already underway by human mathematicians assisted by AI tools.
The takeaway
OpenAI's claim is real in the sense that a formally-verified proof exists and has been published. It is not yet an independently confirmed, prize-worthy resolution of the Millennium Prize problem, and the story behind how it was produced is currently tangled up in a genuine dispute over credit and conduct with another research team. If you're citing this anywhere, the accurate framing is: "OpenAI published an unverified but Lean-checked proof, produced by a large coordinated multi-agent system, addressing a specific formulation of the Navier–Stokes singularity question".
Worth watching over the next few months as outside mathematicians actually dig into the proof.
For more such developer content, click here
Top comments (1)
Good breakdown — especially the "accurate framing" paragraph. Most of the reposts of this story skipped the fact that the proof is Lean-checked but not independently certified, which is a very different claim than "Millennium Prize solved".
What I keep turning over is the cross-pollination step: pulling promising partial results out of agent groups and feeding them back as prompts. That looks like the actual engineering in the story; the rest looks like throwing compute at search. If synthesis was the bottleneck, a 1,000-agent run with better synthesis might land the same place — and that would matter a lot for anyone budgeting multi-agent work.
Would love to see a breakdown of the ~5M messages: averages hide everything. The question for anyone who has actually run agent swarms is whether useful exchanges clustered in a few groups or spread thin across all 10,000.