After AI solved math, it looks like physics is about to be next. GPT-6 Astra achieved a 98% accuracy rate in the latest physics evaluation. A paper jointly published by several scholars in high-energy physics and theoretical physics, titled *TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving*, has created a benchmark dataset for cutting-edge theoretical physics. It includes 50 unpublished, research-grade theoretical physics problems with definitive, verifiable answers. For a single problem, it would take a top expert in the specific subfield about one day to independently derive a solution. If handed to physicists from related fields, it would take even longer. Using the Codex framework, GPT-6 Astra reached a 98% accuracy rate for final answers, and it did so at an extremely low computational cost—average cost per problem was less than $3.