OpenAI used a group of agents powered by a next-generation model, more capable than GPT-6 Astra, to solve the Navier-Stokes Millennium Prize Problem. The system constructed an example of a fluid starting from rest under a smooth external force where a singularity occurs in finite time, solving options C and D of the official problem statement. The company stated it will not claim the $1 million prize from the Clay Mathematics Institute.
The experiment began after OpenAI started training a new internal model on August 28. On September 1, the company deployed groups of interacting agents with access to a cached version of the internet, code execution tools, and the ability to share intermediate results. Approximately 10,000 agents worked on the main task simultaneously, reaching a result on September 5, about 88 hours after launch. GPT-6 Astra was then used for 17 hours to formalize and verify the proof.
During the Navier-Stokes task, agents exchanged approximately 2.7 million messages and generated about 130 billion output tokens. The total mathematical experiment involved 4.9 million messages and approximately 300 billion tokens. Nature magazine described the amount of computing resources used for a single mathematical research project as unprecedented.
A dispute over priority has emerged involving Professor Tristan Buckmaster of NYU and mathematician Levent Alpöge of Harvard and Anthropic. On September 7, they published results on singularities in other fluid dynamics problems. OpenAI admitted it initially thought they had a solution for Navier-Stokes, but claims its proof is substantially different. The company denied accessing specific user data from the researchers' interactions with Codex to reach the result.
On the same day, OpenAI introduced ChatGPT Images 2.5, an image generation and editing model. The new version reduces generation latency by 50% compared to Images 2.0. It introduces the Sketch function, allowing users to draw sketches in the interface as a base for the final image, along with templates for common formats and prompt sharing capabilities.
The company also referenced previous results for its Jalapeño inference chip. In tests, the accelerator showed 1.5–1.9 times more computing work per watt and 1.7–3.6 times lower latency than systems based on Nvidia GB200 and GB300.
