OpenAI makes progress on second Millennium Prize problem

OpenAI has announced that it achieved substantial progress on a second Millennium Prize math problem, according to a statement reported by The New York Times. This development follows a previous breakthrough where the company deployed an unreleased internal model, described as significantly more capable than GPT-6 Astra, to tackle the Navier-Stokes existence and smoothness problem. During that initial milestone, OpenAI coordinated approximately 10,000 agents to run in parallel for about 88 hours, ultimately generating a proposed solution and a formal proof written in the Lean theorem prover.

While OpenAI has not officially named the second Millennium Prize problem it has targeted, industry rumors reported by researcher Andrew Curran point to the Hodge Conjecture. Curran also speculated that the unreleased model driving this latest research might be named Aeon, though OpenAI has not confirmed either detail. The company is currently determining how to share its latest mathematical findings with the broader scientific community.

For AI practitioners and researchers, this achievement highlights the growing viability of massive multi-agent systems for complex reasoning tasks. By throwing immense computing power at problems with checkable outcomes, OpenAI demonstrated how thousands of agents can explore distinct mathematical paths in parallel. The integration of formal verification tools like Lean allows the system to automatically test and verify proposed steps, turning raw compute into a highly structured search for truth.

If these results withstand rigorous peer review from outside mathematicians, it could validate a new paradigm for scientific discovery. Instead of relying solely on human intuition, researchers can leverage agentic networks to automate the tedious process of proof exploration. This shift suggests that verified research output and systematic problem-solving capabilities will increasingly replace traditional static benchmarks as the ultimate measure of frontier AI models.