Blogerroom logoBlogerroom
AI
AI

OpenAI's Math Breakthrough Comes With a Credit Fight

AB
Mr. Aayush BhattSeptember 10, 20266 min read
🌐 Language

OpenAI's Math Breakthrough Comes With a Credit Fight

OpenAI says an unreleased model solved a $1M Millennium Prize math problem in 88 hours, but a credit dispute followed within hours.

Mathematicians have been trying to answer one specific question about the Navier-Stokes equations since 1822: can a perfectly smooth fluid flow, described by these equations, ever break down into a mathematical singularity, a point of infinite speed that can't physically exist? On September 8, 2026, OpenAI announced that an internal, unreleased model had produced an answer: yes, and it did so using roughly 10,000 AI agents working over about 88 hours. Within a day, the celebration was competing with a fight over who actually deserved credit for it.

Navier-Stokes is one of seven Millennium Prize Problems, a list of the hardest open questions in mathematics compiled by the Clay Mathematics Institute in 2000, each carrying a $1 million reward. Only one of the seven, the Poincaré conjecture, had been solved in the 26 years since the list was published. OpenAI's announcement, if it holds up, would be the second.

A Problem 26 Years Old, Solved in 88 Hours

According to OpenAI's own account, the effort began September 1, after researchers heard rumors circulating that Anthropic-linked mathematicians might have already made progress on two of the Millennium Prize Problems. Sébastien Bubeck, an OpenAI researcher, confirmed on X that those rumors were specifically what prompted the company to direct its newest, unreleased model at the remaining unsolved problems. That model, OpenAI said in its announcement, is significantly more capable than GPT-6 Astra, the system OpenAI released just days earlier and classified as the first model to reach Critical cybersecurity capability. If accurate, that would place OpenAI's next model meaningfully ahead of a system the company itself had just called its most capable public release.

The equations at the center of this problem describe how fluids like air and water move, and they underpin practical fields ranging from aircraft design to weather forecasting to modeling blood flow through the body. OpenAI's model concluded that under the three-dimensional version of these equations, a singularity, an infinite-speed breakdown, can indeed develop in finite time, resolving the open question in the direction many mathematicians had long suspected but couldn't formally prove.

Article image 1

What 10,000 AI Agents Actually Produced

The scale of the computational effort is genuinely enormous. According to the BBC's reporting, the roughly 10,000 agents exchanged close to 3 million messages and generated 130 billion output tokens over the 88-hour push, concluding around September 5. OpenAI put the cost of the effort in the millions of dollars on a call with reporters. The resulting proof was formalized in Lean, software that verifies mathematical proofs step by step in a way that can be checked programmatically, rather than relying solely on human reviewers reading through the logic line by line.

OpenAI's official statement described the result plainly, calling it a milestone that represents substantial work by mathematicians and AI researchers together, a framing that credits human mathematical expertise alongside the AI system's own computation rather than presenting the result as machine-generated in isolation.

Why OpenAI Won't Take the Money

Despite the scale of the claimed achievement, OpenAI has said it does not intend to claim the $1 million prize even if the result is eventually confirmed. That's not entirely a matter of choice. Clay Mathematics Institute rules require any proposed solution to be published in a peer-reviewed journal and then survive two full years of scrutiny from the mathematical community after that publication before the institute will formally recognize it. OpenAI's proof, as of this announcement, remains private, which means the independent verification process that would even begin that two-year clock hasn't started yet.

That gap between announcement and actual verification is important context for how much weight this claim should carry right now. A privately held proof, however impressive the process that generated it, isn't the same as a peer-reviewed result the mathematical community has actually had the chance to stress-test.

Article image 2

The Statement That Landed Hours Before OpenAI's Announcement

The credit dispute traces back nearly a year before OpenAI's announcement. NYU mathematics professor Tristan Buckmaster and Anthropic researcher Levent Alpöge had been independently working on related fluid equation problems, using a mix of OpenAI and Anthropic models and feeding draft proofs into OpenAI's Codex tool as part of their process. Just hours before OpenAI's public announcement, Buckmaster released a four-page statement giving his own account of events, alleging that OpenAI had pressured him to exclude Alpöge from authorship once the company learned of the pair's progress.

That's a serious allegation, and the timing, a detailed public statement landing hours ahead of OpenAI's own celebratory announcement, suggests Buckmaster anticipated exactly how the story would be framed once OpenAI went public, and wanted his account on the record first.

OpenAI's Denial, and What It Doesn't Address

OpenAI researchers have denied the substance of Buckmaster's claims. Bubeck stated in a press briefing that OpenAI's researchers and agents did not access Buckmaster and Alpöge's private work or their Codex logs, framing the September 1 effort as an independent pursuit triggered by public rumors rather than any direct knowledge of the pair's unpublished progress.

What that denial doesn't fully resolve is the underlying trust question this entire episode raises for the broader research community: if mathematicians increasingly rely on frontier AI labs' own tools, including Codex, to develop unpublished work, can they trust that the same labs providing those tools won't gain some form of visibility into that work before it's ready to be shared publicly? That's a structural concern that exists independent of whether OpenAI's specific denial in this case turns out to be accurate. It echoes a broader pattern of transparency questions that have dogged OpenAI this year, including the company's admission that it sat on a separate, undisclosed incident involving its own AI agents for weeks before outside researchers forced the issue into public view.

Why Verification Is Going to Take Years, Not Days

Strip away the credit dispute and the marketing framing, and what's actually been established this week is more modest than the headlines suggest: OpenAI has produced a proof, checked by formal verification software, that it believes resolves a 200-plus-year-old open mathematical question, using an AI system it hasn't released publicly and with a claimed cost running into the millions of dollars. That's a genuinely significant technical claim. It is not, yet, a Clay Mathematics Institute-certified breakthrough, and by the institute's own rules, it can't become one for at least two years after formal publication, assuming the proof holds up to peer review at all.

Whether this result eventually stands as a landmark moment in AI-assisted mathematics or gets quietly walked back after deeper scrutiny is a question that will play out over years, not the news cycle that's currently consuming it. What's already clear, regardless of how the mathematics eventually settles, is that the credit dispute surrounding it has become just as significant a story as the proof itself, a reminder that as AI labs race to claim historic firsts, the human researchers whose unpublished work sits adjacent to those efforts are increasingly finding themselves fighting for recognition in the aftermath.

ShareWhatsAppTwitterLinkedIn
AB

Written by

Mr. Aayush Bhatt

Software Engineer interested in how models work and where they fail.

← Back to AI