Key Takeaways

  • OpenAI spent roughly $10 million in compute over an unconstrained one-week experiment to produce a mathematical proof for Navier-Stokes.
  • Test-time compute and chain-of-thought methods now allow researchers to deploy teams of agents running 10,000 parallel rollouts on a single problem.
  • Altman predicts that optimization curves will lower the compute expense of re-running the Navier-Stokes proof to one-hundredth of its original cost within twelve months.
  • Software models previously capped out at roughly one dollar of useful inference compute per query before chain-of-thought opened up parallel agent reasoning.

The Shift From Model Training to Test-Time Agent Teams

For years, the rule in artificial intelligence was simple: train a bigger model, deploy it, and collect answers in seconds. If a question was hard, spending extra server time on the response did not help. As Altman noted to Collison: “A few years ago, we forget how fast this has all gone. But a few years ago, it was impossible to throw more than a dollar of compute at a problem and have it help you at all.”

Chain-of-thought reasoning eliminated that ceiling. Once models could generate intermediate steps and evaluate their own logic, test-time compute became the primary vector for solving hard questions. Instead of querying a static system once, engineers can deploy swarms of models that coordinate, critique each other, and explore distinct solution branches simultaneously. “And then, we figured out chain-of-thought and the models could think hard about a problem,” Altman said. “And now... People really started using a lot of compute. And now, you can throw a whole team at something. And you can do that 10,000 times in parallel.”

The Ten-Million-Dollar Proof and the 100x Cost Curve

To test the limits of agent swarms, OpenAI removed the financial brakes. The team let compute budgets run completely unconstrained for seven days on the Navier-Stokes existence and smoothness problem. The experiment worked, yielding a mathematical proof, but it came with an initial compute bill of roughly $10 million.

OpenAI ran the run without knowing the underlying math of agent coordination. “So, when we did the Navier-Stokes proof, we understood absolutely nothing about the scaling laws of teams of agents,” Altman admitted. Yet brute-force exploration establishes the frontier before engineering refines it. Altman expects the price tag to collapse rapidly: “I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done.” What costs eight figures today becomes a hundred-thousand-dollar job next year, and a routine thousand-dollar background task soon after.

Science Belongs in Public Despite the Disruption

Dropping an automated proof onto a centuries-old problem creates severe shockwaves across academia. Professional mathematicians spend decades trying to crack problems like Navier-Stokes. When an automated cluster solves one in seven days, it raises questions about how scientific fields will absorb non-human discoveries.

Collison and Altman discussed whether releasing such proofs risks demoralizing human researchers. Altman took an uncompromising stance on publication: “Humanity deserves fundamental knowledge. And we should not be sitting on a major fundamental science discovery. The world deserves that and gets to know it.” Holding back progress to soften the transition for human practitioners is, in his view, the wrong trade. Scientific truth belongs to the public domain the moment it is verified.

What to Do With This

Audit your product's hardest automated workflow this week. If you currently make a single API call to a frontier model and return the result to the user, rewrite the backend to spin up three parallel agents that check each other's work and iterate before returning the final output. Track whether spending 5x more on test-time compute eliminates your downstream error rate.