GPT-6 Astra solves 49 of 50 unpublished theoretical physics problems
A new benchmark called TheoPhysEval gave five AI agents 50 unpublished, research-level theoretical physics problems. GPT-6 Astra running in OpenAI Codex solved 49 of them at an average cost under 3 dollars per problem, and along the way it showed that several of the expert answer keys were wrong.