OpenAI's GPT-6-Astra solved 49 out of 50 theoretical physics research problems at an average cost of $2.40 per problem using the Codex coding environment. Thirty-five researchers from institutions including Harvard University, the International Centre for Theoretical Sciences, KU Leuven, and Google DeepMind released the TheoPhysEval benchmark to test AI agents on unpublished calculations.
Anthropic's Opus 5.5 placed second by solving 92 percent of the benchmark, while Google's Gemini 3.8 Flash scored 68 percent, and open-weight models DeepSeek-V4.1-Flash and GLM 5.3 reached 54 and 42 percent respectively. The benchmark covers eleven theoretical physics subfields and tests routine research work rather than grand challenges, utilizing problems that a leading human expert typically takes about a day to solve.
The evaluation revealed that internet access had minimal impact on performance, indicating that advanced language models have largely memorized the theoretical physics literature. Follow-up reviews with experts also uncovered minor errors in the reference solutions, and once those were corrected, Astra's answers were correct for all problems. This performance demonstrates that leading AI agents can now execute complex theoretical calculations as accurately as human specialists in specific subfields.