AI Physics Agent Solves 50 Unpublished Problems at Near-Expert Speed

GPT-6-Astra tested 50 theoretical physics problems with 98% accuracy at under $3 per problem, matching expert time estimates while avoiding web search dependency.

GPT-6-Astra, an AI agent built within Codex, solved 50 unpublished theoretical physics problems with 98% average accuracy. Each problem cost under $3 to complete, far below the estimated one day of work by world experts. The system maintained stability even when web search was denied, and initial errors were corrected through expert review to confirm solutions.

This efficiency could accelerate energy and health research by reducing the time and cost to explore complex physics. For example, faster progress on energy storage or medical physics might lower future costs for clean power or targeted therapies. However, the results only apply to problems with verifiable answers, and the expert time benchmark reflects typical human performance—not actual research workflows. The source material doesn’t detail real-world application timelines or scalability beyond this test set. The source is brief and the detail sits with them.

Source: HumanProgress