GPT-6-Astra solves 49 of 50 unpublished theoretical physics problems in a new benchmark

Thirty-five researchers from institutions including the International Centre for Theoretical Sciences, Harvard University, KU Leuven and Google DeepMind have released TheoPhysEval. The benchmark contains 50 unpublished research problems in theoretical physics with answers that can be checked. According to the authors, a leading expert in the relevant subfield would need about a day or less to solve a typical problem without AI help. Specialists from neighboring fields would need much longer. The team tested commercially available AI agents without giving them any human hints. GPT-6-Astra, running in OpenAI's Codex environment, scored highest. The Decoder reports that it found the correct answer to 49 of the 50 problems, or 98 percent, at an average cost of under $3 per problem and just under $116 for the whole benchmark. The benchmark deliberately leaves out major open questions. It focuses on what the authors call "turn-the-crank" problems: routine calculations that make up much of everyday research work. The results suggest that the best agents can match leading specialists at this kind of work. The authors argue that a missing sense of "taste" is one reason AI has not yet produced breakthroughs in the field. In their view, AI still lacks the judgment needed to decide which problems are worth pursuing.