OpenAI says an experimental system powered by a new internal model has produced a proposed solution to the Navier–Stokes existence and smoothness problem—one of mathematics’ seven Millennium Prize Problems.
The result matters beyond the proof itself. OpenAI coordinated roughly 10,000 concurrent agents in specialized groups, giving them access to code execution and research materials. Codex consolidated promising ideas between groups, while GPT‑6 Astra helped formalize and verify the final argument in Lean.
OpenAI reports that the agents reached the result after approximately 88 hours, exchanging 2.7 million messages and generating around 130 billion output tokens for the Navier–Stokes effort alone. The company published both the mathematical paper and machine-checkable formalization, but described the work as a capability milestone rather than claiming the Millennium Prize. Independent mathematical review and formal recognition remain separate processes.
Why it matters for builders
The strongest signal is not simply that a model generated advanced mathematics. It is that a complete agentic research harness—parallel exploration, tool access, communication between specialist groups, synthesis, and formal verification—was able to pursue one difficult objective at unprecedented scale.
This extends the research direction behind OpenAI’s automated research intern from several parallel experiments into a coordinated system operating at a much larger scale.
For developers, this points toward a new architecture for complex AI work:
- many agents exploring competing approaches in parallel;
- orchestration that reallocates resources as evidence emerges;
- synthesis agents that consolidate useful intermediate results;
- verification systems that test outputs instead of trusting fluent answers;
- persistent infrastructure capable of supporting multi-day research runs.
The internal model is not currently available to developers, but OpenAI says it is significantly more capable than GPT‑6 Astra. That makes this both a research announcement and an early capability preview of models still in development.