Altman Confirms: The Next-Generation GPT That Cracked the Navier-Stokes Equations Is About to Be Released

The "mythical beast" with which OpenAI conquered the Millennium Prize Problems is finally about to be let out of its cage!

Today, at the opening ceremony of ADIPEC 2026, a single remark from Altman set the entire internet ablaze—

The "next-generation GPT," which had 10,000 Agents run nonstop for four days and four nights, working together to crack the Millennium Prize Problems, will be released very soon.

Among these Agents there is no "boss" and no hierarchy; they are able to explore different directions, share progress, and learn from one another.

Next, the highlights of the entire speech, point by point:

The true identity of the next-generation GPT is revealed: Altman confirms that the internal model that cracked the NS equations is the next-generation GPT, to be released soon

Ten-thousand-scale Agents with zero human command: 10,000 Agents collaborate without hierarchy, produce a proof in 88 hours, then complete Lean formal verification in another 17 hours

722 manuscripts dominate the academic world: touching on Riemann and Hodge; the Kakeya conjecture conquered in all dimensions, the exponential time hypothesis overturned, and the existence of Yang-Mills proven

Cross-disciplinary creation right after birth: the same set of capabilities is being used to design new materials and new turbines; Altman says it is "only a few months old"

Over the past two months, this mysterious giant beast has successively smashed through the ceilings of fluid mechanics, analytic number theory, computational complexity, and theoretical physics, leaving countless human luminaries sleepless all night.

And under the spotlight in Abu Dhabi, Altman for the first time recreated the chilling scene of those 10,000 Agents evolving by themselves in the black box.

10,000 Agents team up and battle for 88 hours

A stunning reversal arrives at the 50th hour

Asked about the most shocking AI breakthrough moment of the year, Altman on stage unhesitatingly pointed to the mathematical offensive in early September.

Altman recalled: "Earlier this year, we proved the first Millennium Prize Problem, the Navier-Stokes equations—that was one of the most incredible moments I have seen in the history of AI."

During the four days and four nights in which these 10,000 Agents ran wild continuously, there was no human mathematician at the side pointing out directions, nor was any top-level master control node set up to issue orders.

That was accomplished jointly by the 10,000 Agents of the soon-to-be-released next-generation GPT; they coordinated fully autonomously, chose their own exploration directions, shared progress with one another in real time, learned from their peers' ideas, and independently decided what to do next.

In past multi-Agent engineering, developers were accustomed to placing a "foreman" at the top to break down tasks, while the sub-Agents below merely carried out mechanical execution.

But when facing an uncharted territory like the Millennium Prize Problems, where even humans have never drawn a map, if the preset direction of the master control node deviates even slightly, the entire legion will spin its wheels collectively in a dead end.

The next-generation GPT demonstrated a never-before-seen silicon-based organizational form: throw 10,000 Agents with top-tier intelligence into the same compute sandbox and let them work it out among themselves.

Looking through the underlying logs OpenAI previously made public, those 88 hours of battle unfolded like a sci-fi blockbuster.

On September 1, the campaign began, and 10,000 Agents spontaneously split into two routes: one attempted to prove that three-dimensional fluid solutions are always smooth, while the other searched in reverse for counterexamples of singularity blowup.

A full 50 hours passed, and the sandbox was dead silent; thousands of conventional derivation paths all hit walls.

A turning point appeared around the 50th hour. A small detachment of about 100 Agents was the first to open a breach on finite-time blowup for the force-free Euler equations, and immediately synchronized this discovery to all.

The thousands of Agents that had originally been repeatedly trial-and-erroring in other directions abandoned the work in their hands and quickly pivoted to this new route, concentrating their firepower to push it forward.

By the 88th hour (September 5), tens of thousands of Agents jointly twisted out an extremely exquisite "spaghetti-style spiral vortex flow," rigorously proving that under the condition of bounded total kinetic energy the fluid velocity still rushes toward infinity, ending in one stroke a century-old problem that had stood for 90 years.

Afterward, GPT-6 Astra spent another 17 hours completing Lean formalization and verification. For NS alone, OpenAI disclosed consumption of about 2.7 million messages and 130 billion output Tokens.

Ultimately, it was written into a 165-page magnum opus whose sole author was signed as "OpenAI."

From "a ten-thousand-strong siege for 4 days" to "one shot to the head in 3 hours"

The evolution history of a monster only two months old

The 10,000 Agents besieging the NS equations was only the tip of the iceberg. Digging deeper backward along Altman's remarks, an even more astonishing fact surfaced—

The string of heaven-defying breakthroughs that shook the academic world over the past two months were all, behind the scenes, orchestrated by this soon-to-be-released "next-generation GPT."

The first wave (September 8), when it took down the Navier-Stokes equations, it deployed a legion of ten-thousand-scale Agents for a bitter four-day-and-night battle.

By the second wave (October 6), OpenAI directly released 721 independent preprints, conquering more than 100 long-standing unsolved mathematical problems in one go, of which over 300 came with complete Lean 4 machine verification.

Paper No. 003 even substantially advanced the "Landau-Siegel zero problem," which is directly tied to the Riemann hypothesis, prompting Fields Medalist Terence Tao to personally build a database to verify it.

Then came the third wave (October 26), with three bombshells fired in the same night: it used a 59-page paper to end the 109-year-old all-dimensional "Kakeya conjecture," and a 67-page paper to directly disprove the exponential time hypothesis (ETH) that had ruled the computer science world for 30 years.

Also announced the same day were rigorous mathematical existence proofs for four-dimensional pure Yang-Mills theory and QCD, with even Nobel laureate David Gross and the ETH founders personally leading the review.

The most terrifying turning point was pointed out incisively by well-known AI industry observer Andrew Curran.

After checking the technical report, Curran found that by October, whether it was sweeping through more than 100 unsolved mathematical problems, disproving ETH, or settling the all-dimensional Kakeya conjecture, the ten-thousand-scale Agent swarm had not been deployed at all.

All these century-level breakthroughs were accomplished by the model in a single prompt, "one shot, one kill," with each difficult problem consuming on average only about 3 hours of Pro-level deep thinking, and the model itself had only been born in late August this year.

In other words, what last month still required "a massive army besieging it for four days and nights"—a terminal disease of human intellect—had, one month later, become in its eyes a routine operation of "single prompt, done in 3 hours."

Combined with the hard evidence from the past two days that OpenAI Codex's backend had quietly absorbed 19.8% of real traffic under the name "gpt-6.2-astra," and Codex lead Tibo's intensive moves over 28 consecutive days to expand context and rewrite the parallel foundation for sub-Agents, this compute monster had in fact long since entered the countdown to official launch.

Swallowing 500 billion Tokens in a single month

Earth only worries about not having enough power

As this weapon, once used to top the charts on the millennium prize problems, is about to be stuffed into every ordinary developer's computer, humanity's intellectual balance was also completely rewritten overnight.

After hearing the talk, Professor Derya Unutmaz, a well-known immunologist at the Jackson Laboratory in the United States, excitedly posted that the scene of 10,000 Agents autonomously collaborating to crack the millennium prize problems proclaimed the official arrival of the "superintelligence era."

He predicted that as this "next-generation GPT" opens up to academia and the public, humanity is about to see an epic explosion of discoveries in conquering terminal diseases in the life sciences such as cancer, Alzheimer's disease, and reversing human aging.

To corroborate the intensity of this compute storm, Altman also took the stage to reveal for the first time a set of jaw-dropping internal Token bills.

He disclosed that six years ago, the world's number one Token consumer was an internal OpenAI employee, using about 100,000 Tokens per month; today, the global per capita monthly consumption across 8 billion people has already caught up to that 100,000 Tokens from back then.

And now the world's number one "mystery top spender," a single person, consumes as much as a staggering 500 billion Tokens in one month—a 5-million-fold surge in just six years.

An ordinary person chatting could never use even a fraction of 500 billion Tokens in a lifetime; the only thing capable of swallowing such astronomical figures is precisely the round-the-clock, fully automated Agent swarms for scientific research and engineering.

Facing this compute storm, Altman, holding the next-generation GPT, also admitted at the end that as AI begins to take over scientific research and chip engineering, the number one bottleneck before him today comes down to just one word: "energy."

And for developers and scientists around the world, that "internal divine beast" that once used 10,000 Agents to crack the millennium prize problems has already entered the launch countdown—

Before long, this brain that even top scholars cannot hope to match will quietly lie in everyone's chat box.

References:

This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelation, editors: Taozi, Moses, published by 36Kr with authorization.