Claude advances a demanding physics calculation in days

Anthropic says its Claude AI has pushed a demanding particle-physics calculation from eight to nine “loops”, completing the work in several days with little human supervision.

The calculation concerns scattering amplitudes, mathematical tools for predicting how particles interact. Each loop adds another layer of correction, making the work increasingly difficult. This problem involved six particles in a simplified theoretical model used to develop and test methods, rather than directly describe the real world.

Claude used techniques developed by physicist Lance Dixon and his collaborators, applying known physical rules to narrow down possible answers. It completed the calculation by two separate routes, which produced matching results. Each route was estimated to cost users US$1,000 to US$2,000.

Dixon, of SLAC National Accelerator Laboratory and Stanford University, independently verified the result. His team had achieved the previous eight-loop benchmark. The calculation files have also been made publicly available for others to examine.

The achievement is not the discovery of a new physical law. It shows an AI system successfully carrying out an intricate research workflow built on years of human work. The methods still need wider independent scrutiny, and it remains uncertain how far this success can extend to calculations describing real-world physics.