The headline
On September 8, 2026, OpenAI announced that an internal model — described only as "significantly more capable than GPT-6 Astra" — had produced a proof that 3D Navier-Stokes equations can develop a singularity in finite time. That's one of the seven Clay Millennium Prize Problems. Ninety years unsolved. $1M bounty attached, which OpenAI says it doesn't want.
The method: throw roughly 10,000 agents at it in parallel, let them grind for 88 hours, spend "millions of dollars" in compute, and out comes a proof. GPT-6 Astra needed another 17 hours just to verify it.
If that headline stood on its own, it'd be one of the biggest AI-for-science stories of the year. It doesn't stand on its own.
The part that didn't make the blog post
Three days after the announcement, NYU mathematician Tristan Buckmaster published a long statement on his own site. Short version: he and Levent Alpöge — who works at Anthropic — had spent the better part of a year chasing a closely related result using a niche method almost nobody else in the field was working on. They hit a real breakthrough on August 15, verified it by August 22, and spent the following week turning it into something publishable.
Where were they doing that work? In OpenAI's Codex, because that's the tool they used to write and iterate on the proof.
On September 1 — six days before Buckmaster's team was ready to publish — OpenAI kicked off its own Millennium Problems sprint. In their own words: "we heard rumours that two Millennium Prize problems had been resolved... inspired by these rumours... we launched an effort to evaluate [our model] on all open Millennium Prize problems."
Buckmaster asked OpenAI researcher Sébastien Bubeck directly, on a call, whether the model had been trained on or had access to their Codex sessions — sessions that contained every draft of their unpublished work. He says he didn't get a straight answer on training, and got told the model "did not look up user data" on access.
Then, per Buckmaster, OpenAI gave him two options: publish jointly with OpenAI crediting his team, or write up the Navier-Stokes result solo — leaving Alpöge off the author list because he works for a competitor.
He said no to both, and said he'd go public. His account of what came back: "Why would you ruin your career?"
Bubeck's response, posted publicly, walks that quote back and calls it "an extremely poor choice of words" that he "retracted on the spot." He says he never asked for Alpöge's removal, only suggested Buckmaster lead-author a rewrite of OpenAI's own proof, and that OpenAI's team never saw the pair's work before it went public. Altman chimed in on X too: "we did not rush to publish... the team threatened us with unfounded accusations of plagiarism."
OpenAI's own statement threads the needle carefully: no researcher or agent accessed the pair's specific data, "no specific user data was accessed in order to solve this problem," but — and this is the sentence that should worry every engineer using Codex, Cursor, or Copilot for anything sensitive — "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
Read that again. The company that makes the tool cannot rule out that your private, unpublished work fed the model it's now racing you with.
Why this isn't a math story
Nobody's claiming the two proofs are the same — Bubeck says "our proofs differ significantly and even the precise results proved are different in the Euler case." Fine. Buckmaster himself is careful to say "I am not accusing anyone of anything." This isn't a plagiarism trial.
It's a structural problem, and it's the same one every developer using an AI coding tool for proprietary work should already be uneasy about:
The vendor of your tool is also a research competitor with more compute than you'll ever have access to. You put unpublished, uncommitted, half-finished work into Codex because that's what it's for. The vendor's usage terms say enterprise data isn't trained on by default, personal accounts can opt out. But "by default" and "can opt out" are policy promises, not architectural guarantees — and this incident is the first time we've watched, in public, what happens when the policy gets stress-tested by a plausible reason to want the data.
Nine months ago this would've been a hypothetical raised in a security review. Now it's a documented sequence of events: rumor of a competitor's breakthrough → six days later, a sprint targeting that exact problem → a result that took the exact adjacent path.
Could it be coincidence? Genuinely, maybe. Convergent proof strategies happen in math. But "genuinely, maybe" is not the bar you want your tooling vendor held to when your competitive advantage is sitting in their session logs.
The technical claim, on its own merits
Set the drama aside for a second, because the underlying capability claim is worth taking seriously on its own.
- 10,000 agents working "somewhat concurrently" for 88 hours is not a chat completion. It's an orchestration problem — task decomposition, partial-result verification, backtracking when a branch of the proof search dead-ends, and some mechanism for merging independently-generated lemmas into a coherent argument. OpenAI hasn't published the orchestration architecture, which is its own tell.
- 17 hours of verification by a different model (GPT-6 Astra, not the system that produced the proof) is the closest thing to independent replication we've gotten from a lab announcing its own result. It's not independent. It's still owned by the people grading their own homework.
- No paper. No arXiv preprint. No third-party mathematician has confirmed the proof is correct as of this writing — only that OpenAI's summary of it is internally consistent enough to publish a blog post about. "Solved" is doing a lot of work in every headline including mine.
Compare this to how actual math breakthroughs get treated: quietly, for months, while independent experts tear the proof apart looking for the hole. OpenAI's version compressed that into a press call. That's a marketing choice, and it's the choice that manufactured this entire mess — a real result, if it holds, wrapped in a rollout designed for virality instead of verification.
What to actually take away from this
If you're an engineer putting real work into any AI coding tool — Codex, Copilot, Cursor, Claude Code, doesn't matter which — this is the concrete scenario your threat model was missing:
- "We don't train on your data by default" is a setting, not a wall. Settings get toggled, inherited wrong, or quietly reinterpreted under "de-identified aggregate improvement" language that every provider's ToS already contains.
- The moment your provider has a commercial incentive to want what you're building, the paper trail matters. Anthropic vs. OpenAI is the visible case because both are frontier labs competing on the same benchmarks. Your startup's roadmap sitting in a coding assistant's context window has the exact same exposure, just with nobody watching.
- Air-gap what you can't afford to lose. If you're doing genuinely novel, competitively sensitive work — research, a proof, an unpublished algorithm, a product nobody's shipped yet — treat the AI tool as a public collaborator until proven otherwise, not a private notebook.
OpenAI may well have a real, independently-verifiable proof of a 90-year-old problem. If so, that's a legitimately enormous result and worth celebrating once it survives peer review. But the way it got here — a rumor, a sprint, a call where someone allegedly asked "why would you ruin your career," and a policy statement that can't rule out training on the other guy's private sessions — is the story every engineer using these tools needs to sit with longer than the headline.
The math might check out. The trust model already didn't.
Top comments (0)