OpenAI reportedly solved a Millennium Prize Problem using 10,000 agents, researcher says

Brown described how performance improves when multiple agents work together: four agents roughly halve the time at double the cost, while 16 agents give diminishing returns depending on the task. He noted that math problems appear highly parallelizable, whereas tasks like novel writing are not. The conversation also touched on the rapid progress in AI‑driven mathematics, with models moving from high‑school benchmarks in 2024 to International Math Olympiad gold in 2025, and now solving open problems faster than expected. Brown warned that the science of scaling to tens of thousands of agents is still nascent, and that measuring the exact speedup from 10,000 agents remains difficult.