OpenAI says it has made progress on a math problem that has puzzled researchers for 90 years. The company announced on Tuesday that a new AI model helped solve part of the Navier-Stokes existence and smoothness problem.
The Navier-Stokes equations describe how liquids and gases move. They sit at the center of turbulence, a phenomenon scientists still don't fully understand.
OpenAI says it used roughly 10,000 AI agents working together to reach the result. The bots completed the task in about 88 hours.
How OpenAI Reached The Result
At the end of August, OpenAI began training a new AI model. The company says the model quickly showed strong skill at math.
Researchers decided to test it on some of the toughest unsolved math problems. On September 1, OpenAI heard rumors that two Millennium Prize problems had already been solved by someone else.
That pushed the company to move fast. It set thousands of AI bots to work on the remaining problems using the new model.
By September 5, the bots had produced a solution. OpenAI says the agents exchanged close to 3 million messages during the process.
The company estimates the work used 130 billion output tokens. Based on OpenAI's own pricing, that effort would have cost around $10 million.
The solution addressed two of the four statements required for the full Millennium Prize proof. The prize is worth $1 million and is overseen by the Clay Mathematics Institute.
OpenAI has not claimed the prize. The company says its goal was to show how far its AI models have advanced, not to collect the reward.
A Dispute Over Timing
The announcement has drawn pushback from outside researchers. Tristan Buckmaster, a math professor at New York University, says he was already working on the same problem.
Buckmaster worked alongside Levent Alpöge, a mathematician at OpenAI rival Anthropic. The pair used OpenAI's coding tool Codex during their research.
Buckmaster says he learned on September 3 that details of his team's progress had reached OpenAI. He claims OpenAI did not start its own work on the problem until after that point.
OpenAI published its results a few hours after Buckmaster made his claims public. The company called the timing a coincidence and described Buckmaster and Alpöge's work as separate progress on the same problem.
OpenAI says it did not view any of the outside team's work before it was released publicly. The company also says no user data was accessed as part of its process.
"While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models," OpenAI said. The company added that its proof and its results differ from the other team's work.
Buckmaster has said he has not yet reviewed OpenAI's full proof. He says he felt he needed to share his side of events regardless.
The new AI model used for the project remains internal to OpenAI. The company has not said when, or if, it plans to release the tool publicly.
