Advanced AI systems that devote substantial computing resources at inference time rather than solely during training are emerging as powerful engines for scientific discovery. OpenAI’s president has highlighted this shift, noting that models capable of extended “thinking” or multi-step reasoning during testing are unlocking progress on long-standing problems that previously resisted human effort alone.
A striking example came in September 2026, when OpenAI announced a solution to a key formulation of the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems in mathematics. An internal next-generation model coordinated roughly 10,000 AI agents that worked for about 88 hours. The agents produced an analytical proof showing that a smooth fluid under a smooth external force can develop a finite-time singularity while maintaining finite energy. The result was then formalized and machine-verified in the Lean proof assistant. While the Clay Mathematics Institute has not yet officially accepted the solution and some aspects remain under discussion among mathematicians, the demonstration illustrated the new power of large-scale agent-based search combined with formal verification.
Complementing this high-profile mathematical advance, Anthropic released a detailed study of how researchers actually use its Claude models. Analyzing more than 200,000 scientific sessions, the company found that AI currently handles a median of 38 percent of the time researchers spend on typical studies primarily in data analysis and writing. In fields such as computer science and statistics, widespread adoption of existing capabilities could roughly double research output. In more speculative scenarios where AI manages nearly all non-physical tasks, productivity gains could reach far higher levels, though laboratory and field sciences remain constrained by physical experimentation.
Industry leaders see even broader implications. Box CEO Aaron Levie has argued that “agent swarms” large numbers of coordinated AI agents working in parallel transform many research and discovery problems into ones limited mainly by available compute. Tasks once bottlenecked by human planning, review, and coordination can now be scaled dramatically, opening the door to higher volumes of exploration in science, cybersecurity, materials research, and beyond.
Despite the momentum, significant limits remain. These systems still require enormous computational resources, careful human oversight to validate results and set meaningful goals, and rigorous external scrutiny especially for claims involving foundational mathematical problems. As models grow more capable at long-horizon reasoning and collaboration, the combination of extra test-time compute and multi-agent architectures appears poised to accelerate scientific progress in ways that were difficult to imagine only a few years ago.
