OpenAI Claims a 90-Year-Old Math Problem Solved by Internal AI

OpenAI announced on Tuesday that an internal AI model, more capable than its newly released GPT-6 Astra, has produced a solution to the Navier-Stokes problem, a fluid-dynamics equation that has stumped mathematicians for around 90 years. The system used 10,000 concurrent agents to reach the result. Navier-Stokes is one of seven Millennium Prize Problems, a set of math challenges each worth $1 million to whoever solves them. The claim, first reported by The New York Times and Wired, is already drawing intense scrutiny.

What happened

Detail	Fact
Problem solved (claimed)	Navier-Stokes equations (fluid and gas flow)
Years unsolved	~90 years
Model used	Internal AI, described as more powerful than GPT-6 Astra
Agents running concurrently	10,000
Training start date	August 28th
Prize value	$1 million (one of seven Millennium Prize Problems)

OpenAI published a blog post Tuesday claiming its internal AI model found a solution to the Navier-Stokes problem, a set of equations that describe how liquids and gases move. Getting these equations to work under all conditions has defeated mathematicians for about nine decades, and a verified proof would qualify for a $1 million Millennium Prize from the Clay Mathematics Institute.

According to OpenAI, training on this internal model began August 28th, and the system has “exhibited unprecedented” capabilities. The company says the result was reached using 10,000 agents working at the same time, not a single model pass. First reports came from The New York Times and Wired before OpenAI published its own post.

Why it matters

The Navier-Stokes equations are not an abstract curiosity. They underpin weather forecasting, aircraft design, engine efficiency, and climate modeling. A true proof of their global smooth solutions (the specific version the prize targets) would settle a foundational question in physics and mathematics at the same time.

For the AI industry, the claim matters for a different reason. If an AI system can credibly crack a Millennium Prize Problem, it signals that large-scale multi-agent architectures (systems where thousands of AI instances work on sub-problems in parallel) can reach genuinely novel scientific results, not just summarize existing knowledge.

It also raises the stakes for OpenAI commercially. A verified Millennium Prize win would be the most concrete demonstration yet that frontier AI earns its price tag on hard, real-world problems. You can see how teams are already starting to use AI for complex technical work in our coverage of Cognition’s $2B raise around AI coding.

Why there is already drama

Mathematical proofs require independent verification. No external mathematician or institution has yet confirmed the result. The Clay Mathematics Institute, which administers the prizes, has its own review process, and past claimed solutions to Millennium Problems have collapsed under peer scrutiny.

The use of an AI model to generate a proof also raises methodological questions. A formal proof must be logically complete and checkable step by step. Whether a multi-agent AI system can produce output that meets that standard, rather than a plausible-looking approximation, is exactly what reviewers will be asked to judge.

Our take

We are skeptical, and you should be too, at least for now. OpenAI has every incentive to announce this result boldly and every commercial reason to do so ahead of competitor moves. The math community has seen dramatic claimed solutions to hard problems before, and most do not survive a careful read.

That said, the architecture here is worth watching regardless of the outcome. Ten thousand concurrent agents tackling sub-problems in parallel is a real shift in how AI is being deployed for research. Even if the Navier-Stokes claim is partially wrong or incomplete, the approach points toward how AI tools will be applied to science and engineering problems over the next few years.

For business operators, the immediate takeaway is not “AI solved math.” It is that multi-agent AI systems are now being aimed at problems that were previously considered beyond any automated approach. If your business has hard optimization, simulation, or modeling problems, this is the space to watch. Our AI integration work often starts with exactly this question: where does a hard internal problem become tractable when you throw coordinated AI agents at it?

What to do about it

1. Watch for independent verification from academic mathematicians or the Clay Mathematics Institute before treating this as settled.
2. Read OpenAI’s original blog post and the Wired and New York Times coverage side by side to compare what is claimed versus what is independently confirmed.
3. Note the August 28th training date and track any follow-up benchmarks OpenAI releases for this internal model.
4. If your team uses AI for technical or scientific work, start asking vendors how their systems handle multi-agent parallelism, because that architecture is moving fast.

Verify first, then be impressed: a claimed proof and a verified proof are two very different things in mathematics.

Source: The Verge · AI