AI OpenAI Solved the Navier-Stokes Problem, Which Mathematicians Could Not Solve for a Quarter Century

cryptonews.ruPublished on 2026-09-09Last updated on 2026-09-09

Abstract

On September 8, 2026, OpenAI announced that its AI system had solved the problem of the existence and smoothness of solutions to the Navier-Stokes equations, one of the seven Millennium Prize Problems set by the Clay Institute in 2000. The problem, unresolved for over a quarter-century, asks whether smooth, initially calm fluid flows can develop a singularity—infinite velocity in finite time—under a smooth external force. OpenAI's system proved that such a singularity can occur, with the total energy of the system remaining finite. This corresponds to two of the possible outcomes outlined by the Clay Institute. The work was performed by an unreleased, advanced model. Starting on September 1, a coordinated group of approximately 10,000 AI agents worked for about 88 hours, exchanging millions of messages and generating roughly 130 billion tokens of text. The solution was formalized in the Lean proof assistant for verification. OpenAI stated it does not intend to claim the $1 million prize. On the same day, mathematicians Tristan Buckmaster and Levent Alpöge independently published proofs of finite-time singularities for related fluid equations. They raised questions about whether their work in OpenAI's Codex sessions might have influenced the company's model, but OpenAI denied accessing their specific data, noting the proofs differ in key aspects. The mathematical community will now review both results, a process expected to take months. If validated, it would mark a mileston...

On September 8, 2026, OpenAI announced that its artificial intelligence-based system has solved the problem of the existence and smoothness of solutions to the Navier-Stokes equations. This is one of the seven "Millennium Problems" — a list of the most important unsolved problems in mathematics compiled in the year 2000 by the American Clay Mathematics Institute. A $1 million prize is offered for solving each of them, but this particular problem was considered one of the most stubborn: more than twenty years had passed since the question was posed, with no complete answer.

To understand what this is about, it's worth explaining what the Navier-Stokes equations are. This is a system of formulas that describes the motion of liquids and gases — from water flowing in a pipe to air currents around an airplane wing. The problem is that mathematicians have so far been unable to prove one seemingly simple thing: do solutions to these equations always remain "smooth," meaning predictable and without sharp jumps, or under certain conditions can a fluid accelerate to infinite speed in a finite time — the so-called singularity or "blowup" point of the solution.

OpenAI stated that its system proved that a smooth fluid, initially at rest, when subjected to a smooth external force, can indeed develop such a singularity in finite time — its velocity tends to infinity, while the total energy of the system remains finite. This corresponds to two of several possible answer options to the problem, labeled with the letters "C" and "D" in the official Clay Institute formulation.

Visually, the process is described as follows: a vortex forms in the fluid, which twists inward on itself while simultaneously stretching into an increasingly thin thread — OpenAI compared this to stretching spaghetti. The central region of the vortex compresses to an infinitesimal point, and the rotational speed within it increases so rapidly that in the limit it becomes infinite. At the same time, the key mathematical subtlety of the proof is to show that the energy required for such a process does not escape to infinity along with the velocity but remains a finite number.

In addition to the proof itself, the company published its formalization in the Lean language. This is a specialized programming language and proof assistant system for verifying mathematical proofs: with it, one can verify step by step, line by line, that there are no missing assumptions or hidden errors in the proof's logic — meaning that verifying the proof can be partially performed by a computer, not just by a human.

How the system arrived at the solution

The work was performed by an internal, not-yet-released OpenAI model that, according to the company, significantly surpasses its current flagship model GPT-6 Astra in capabilities. Training of this new model began on August 28, 2026.

On September 1, 2026, after rumors emerged that the company was close to solving two Millennium Problems at once, OpenAI launched a system of coordinating AI agents — essentially many parallel working copies of the model that could exchange intermediate results, access a saved copy of the internet, and independently write and execute program code to verify their computations.

The group working specifically on the Navier-Stokes equations consisted of roughly 10,000 such agents working simultaneously. The solution was obtained on September 5, 2026 — after about 88 hours, or just over three and a half days of continuous work. An additional 17 hours were spent on further formalization and verification of the result using the GPT-6 Astra model.

The scale of the computations can be assessed by the number of "negotiations" between agents: during the work, they exchanged 2.7 million text messages and in total generated about 130 billion tokens — fragments of text from which language model responses are composed. For comparison, this is many orders of magnitude greater than the volume a human would require to write such a proof manually.

Even earlier, before the main work, about 100 agents managed in approximately 50 hours to solve a related, simpler problem — they proved the regularity (i.e., absence of singularities) of solutions to the Euler equations in the so-called unforced case, when no external force acts on the fluid. The Euler equations are a simplified version of the Navier-Stokes equations that does not account for the internal friction of the fluid (viscosity).

Although the Clay Institute prize can formally be awarded for solving this problem, OpenAI stated that it does not intend to claim the $1 million award and published the result primarily to show how rapidly the capabilities of its models in solving complex scientific problems are growing. An official post by the company on social network X on the same day confirmed these details.

Why scientists from another company got involved

On the same day, September 8, 2026, another interesting coincidence occurred: independently of OpenAI, Professor of Mathematics at New York University Tristan Buckmaster and researcher from the competing company Anthropic Levent Alpöge published their results. They also proved the existence of a finite-time singularity under a smooth external force, but for a different set of equations:

  • equations of incompressible porous media — they describe how fluid seeps through porous material, such as soil;

  • the Boussinesq system — a simplified model used to describe convection processes, i.e., the mixing of a fluid or gas when heated;

  • three-dimensional incompressible Euler equations — that same simplified, "inviscid" version of the fluid motion equations.

Buckmaster stated that in their work they relied on the Codex model (another OpenAI product designed for writing and analyzing code) and other models, and stored drafts in Codex working sessions. After this, the scientists contacted OpenAI with a question about whether the company had gained access to their developments through these sessions.

OpenAI responded that it had not seen Buckmaster and Alpöge's work before its public release and that specific user data was not used to obtain its own solution. At the same time, the company clarified: it cannot completely rule out that anonymized data collected during the use of its products somehow influenced the general training and improvement of the models. An additional argument in favor of the independence of the two works is that the proofs themselves differ: in particular, Buckmaster and Alpöge's version of the Euler equations is forced (i.e., accounts for external force), whereas OpenAI's is unforced.

What will happen with the verification of the result

Neither the mathematical community nor the Clay Institute itself had issued official recognition of OpenAI's result at the time of publication. Formalizing the proof in the Lean language simplifies verification but does not completely replace it: experts must ascertain that all steps of the proof indeed correspond to the conditions of the problem formulated by the institute more than twenty years ago. Estimates suggest such verification could take months.

If the result is ultimately confirmed by independent specialists, this will become one of the first cases where a problem of this level of complexity was solved not by a team of mathematicians, but by a system of thousands of AI agents working in parallel over several days. For now, both proofs — from OpenAI and from Buckmaster and Alpöge — remain subjects of study by the scientific community.

AI Opinion

From the perspective of machine data analysis, the situation looks like a continuation of a long-standing mathematical narrative, not its beginning. Already in 2014, Terence Tao constructed an "averaged" version of the Navier-Stokes equations, in which a smooth solution still accelerated to infinity in finite time — at the cost of abandoning part of the structure of the actual equations. That result served as a guide for a decade: it showed precisely where any proof attempt should aim and explained why the problem remained unsolved for so long. The current OpenAI announcement claims to go a step further — the singularity is derived for the full, un-truncated system. The technical aspect remains off-screen: formalization in Lean addresses logical gaps but does not address the question of the correctness of the physical model of the external force applied to the fluid. Will the singularity remain physically realizable, or is it a consequence of the chosen class of external influences?

end-content

Related Questions

QWhat did OpenAI claim to have solved regarding the Navier-Stokes equations, and why is this significant?

AOpenAI claimed that its AI system solved the problem of existence and smoothness of solutions to the Navier-Stokes equations. This is significant because it is one of the seven "Millennium Prize Problems" established by the Clay Mathematics Institute in 2000, with a $1 million prize for its solution. The problem, concerning whether fluid flow described by these equations can develop singularities (infinite velocity in finite time) under certain conditions, had remained unsolved for over two decades.

QHow did OpenAI's AI system work to find the proof, and over what timeframe?

AOpenAI's yet-unreleased internal model, surpassing GPT-6 Astra, coordinated a swarm of approximately 10,000 AI agents. These agents worked in parallel, exchanging intermediate results, accessing a saved copy of the internet, and writing/executing code to verify computations. The proof for the Navier-Stokes problem was obtained on September 5, 2026, after about 88 hours (roughly 3.5 days) of continuous work, followed by 17 hours for formalization and verification using GPT-6 Astra. The agents exchanged 2.7 million text messages and generated about 130 billion tokens during the process.

QWhat specific result did OpenAI's proof demonstrate about fluid behavior?

AOpenAI's proof demonstrated that a smooth fluid initially at rest, under the application of a smooth external force, can develop a singularity (where its velocity becomes infinite) within a finite time, while the total energy of the system remains finite. This corresponds to two of the possible answer options (labeled 'C' and 'D') in the Clay Institute's official problem statement. The process is visualized as a vortex stretching into an infinitely thin filament, with its core region compressing to a point and rotational speed growing to infinity.

QWhat was the independent contribution from mathematicians Tristan Buckmaster and Levent Alpöge, and how did it relate to OpenAI's work?

AIndependently of OpenAI, mathematicians Tristan Buckmaster and Levent Alpöge published a proof on the same day showing finite-time singularity formation under a smooth external force, but for a different set of equations: the incompressible porous medium equation, the Boussinesq system, and the three-dimensional incompressible Euler equations (the inviscid version of Navier-Stokes). They used tools like OpenAI's Codex but clarified their work was separate. A key difference is that their Euler equations result is for the forced case (with an external force), while OpenAI's related prior work proved regularity for the unforced Euler equations.

QWhat is the current status and expected next steps for verifying OpenAI's claimed proof?

AAs of the article's publication, the mathematical community and the Clay Mathematics Institute have not yet officially recognized OpenAI's result. The proof's formalization in the Lean programming language aids verification by allowing step-by-step logical checking by computer, but experts must still verify that all steps correctly address the original problem's conditions. This independent verification process is expected to take several months. If confirmed, it would mark a milestone as a problem of this complexity was solved primarily by a coordinated swarm of AI agents.

Related Reads

GPT-6.1 Sol Participates in Live Brain Surgery! Three Rounds of Localization, Each More Accurate Than the Last

Exin Therapeutics CEO Gabriel Ocana-Santero reported an experiment in which GPT-6.1 Sol participated in three rounds of real mouse brain surgery positioning. The task involved locating Bregma and Lambda, two key skull landmarks used to calibrate stereotaxic coordinates for accurate targeting of brain regions. After each round, a professional surgeon corrected the model's positioning results, and the model retained its images, measurements, and expert corrections for reference in subsequent rounds. By the third round, GPT-6.1 Sol reduced landmark search time from 169 to 86 seconds, required corrections from four to two, and cut positioning error from 0.6 mm to 0.3 mm. Ocana-Santero described this as in-context learning from real-world experience. The team did not disclose how the experience records are integrated into the model or whether underlying weights are updated, so it cannot be described as autonomous training or parameter-level self-evolution. The sample is only three operations, far from proving stable surgical capability, but it demonstrates AI accumulating experience from real observations and expert feedback. This follows earlier Exin work in which AI selected gene therapy vectors, chose injection targets and coordinates, and controlled a stereotaxic device in a mouse experiment, with a licensed veterinarian supervising and safety limits in place. In a further experiment, Claude Sonnet 5, GPT-6.1 Sol, and Grok 4.7 chose different brain regions and treatment strategies for the same scientific goal, suggesting models may reason toward different experimental paths.

marsbit2h ago

GPT-6.1 Sol Participates in Live Brain Surgery! Three Rounds of Localization, Each More Accurate Than the Last

marsbit2h ago

From Drones to Nuclear Reactors: What Has Paradigm Invested in Since Leaving the Crypto World?

Paradigm, a venture capital firm known for crypto investments like Coinbase and Uniswap, is rapidly expanding into hard tech. In November, it co-led a $1.5 billion funding round for Zipline, a US drone delivery company now valued at $21.5 billion—nearly triple its valuation from January. Zipline, founded in 2014, began delivering medical supplies in Rwanda and now operates a large-scale automated logistics network, having completed 150 million miles of commercial autonomous flights. Walmart plans to add 300 new delivery sites, and Uber aims to scale their partnership to one million daily orders by 2029. Paradigm's broader portfolio now spans AI, robotics, manufacturing, semiconductors, space defense, and nuclear energy. Investments include open-source AI lab Nous Research, math-reasoning AI firm Harmonic, AI data company Idler, and GPU platform Andromeda. It also backs SendCutSend (metal parts manufacturing), Fab2 (compact semiconductor factories), True Anomaly (space defense), and Antares (micro nuclear reactors). In July, Paradigm launched a $12 billion new fund for frontier tech. This shift reflects a broader VC trend toward AI and physical technology, as crypto's path to sustained commercial revenue remains uncertain. Paradigm's bets focus on companies solving concrete industrial problems with existing markets, aiming to build a hard-tech network that transcends crypto cycles.

marsbit2h ago

From Drones to Nuclear Reactors: What Has Paradigm Invested in Since Leaving the Crypto World?

marsbit2h ago

Nomura: "Blue Wave" Hits US Data Centers, China's AI Benefits Instead

Nomura argues that the Democratic "blue wave" in the US midterms could paradoxically benefit China's AI economy. Democratic victories in key data-center states will likely tighten regulation, shift grid upgrade costs onto developers, and lengthen approval timelines, slowing US data-center expansion. Market reaction was sharp: the Nasdaq fell 5.5% in a single day, the SOX dropped nearly 15% over two days, Nvidia fell about 12%, and Micron about 24%. Because US cloud giants had been hoarding global GPU, HBM, and advanced logic chip capacity—driving black-market premiums of 200–300% at Asian logistics hubs—a US capex slowdown could collapse those speculative premiums, letting Chinese AI firms, cloud providers, and research institutions obtain GPUs, HBM modules, and advanced packaging materials more cheaply through intermediary trade hubs. China is also building a state-led national AI data-center network, with roughly RMB 2 trillion planned under the 15th Five-Year Plan to integrate regional computing by 2028, supported by low-cost renewable power in western hubs under the "East Data, West Computing" program. Morgan Stanley, meanwhile, does not see the blue wave as a real easing of US-China tensions; it expects strategic competition to persist but sees greater policy predictability. It prefers China's "self-sufficiency" AI trade—supported by both supply and demand under national policy priorities—over export chains tied to global AI capex. It also notes more balanced opportunities in Chinese equities, with valuations near historic lows.

marsbit3h ago

Nomura: "Blue Wave" Hits US Data Centers, China's AI Benefits Instead

marsbit3h ago

After GPT-6, Embodied Large Models Are Trapped in Their Own Territory! Next Year's Self-Research Budget Is Probably Up in the Air

The article discusses how the release of GPT-6 Astra (and the imminent GPT-6.2 Astra) has disrupted the embodied AI/robotics industry. General-purpose large models are now encroaching on capabilities that embodied AI companies previously developed in-house—such as in-context learning, out-of-distribution generalization, error correction, and open-vocabulary manipulation—often without task-specific post-training. On benchmarks like RoboDojo, teams combining general models with robot-use-agent techniques saw scores jump from 20–30 to 50–60, prompting many to question whether self-developing embodied foundation models still makes sense. Some companies have begun cutting algorithm roles. However, the article notes GPT-6 Astra still struggles with complex physical contact, force control, object deformation, and high-precision manipulation, and its cloud API inference latency is high (though this is not seen as a fundamental barrier). Experts argue that general models' current weaknesses shouldn't be treated as durable moats for embodied companies, since OpenAI is already hiring robotics talent and future models may absorb more robot operation data. The core argument is economic: embodied companies' opportunity lies not in building smarter models but in delivering better economic value. Stable, standardized industrial tasks may be better served by lightweight, specialized models or rule-based strategies, while open-ended tasks (e.g., home service robots) may benefit from general models. Ultimately, companies need to re-prioritize their R&D directions—leveraging general models for language understanding and task planning while focusing in-house efforts on physical interaction, embodiment-specific control, real-world adaptation, and deployment efficiency. The article concludes that the industry is moving past the demo-and-burn-investment stage, and the elimination round among the "hundred robot wars" has begun.

marsbit3h ago

After GPT-6, Embodied Large Models Are Trapped in Their Own Territory! Next Year's Self-Research Budget Is Probably Up in the Air

marsbit3h ago

Trading

Spot
活动图片