GPT-6-Astra 解決 50 道物理研究問題中的 49 道，每道成本 2.40 美元

thirty-five researchers from institutions including icts, harvard university, ku leuven, and google deepmind created the theophysEval benchmark to test ai agents on theoretical physics calculations. gpt-6-astra, running in openai's codex coding environment, solved 49 of 50 problems or 98 percent at an average cost of $2.40 per problem and $116 for the whole benchmark. the benchmark focuses on routine research work, not grand challenges, and tests ai agents' ability to apply known methods to slightly modified questions. this achievement shows ai agents can match leading specialists in some subfields of theoretical physics, but also highlights that a missing sense of 'taste' is one reason breakthroughs haven't followed yet.