GPT-6-Astra Matches Expert-Level Performance in Theoretical Physics

AI in the Service of Theoretical Physics

A major breakthrough has occurred in theoretical physics. A new study, involving researchers from 35 scientific institutions including Harvard University and Google DeepMind, has demonstrated that modern AI agents can solve complex physics problems with the same accuracy as leading experts in the field. The project utilized a new benchmark called TheoPhysEval to test AI capabilities across 50 previously unpublished scientific problems covering cosmology, particle physics, biological processes, and fluid dynamics.

OpenAI’s GPT-6-Astra delivered the best performance in the testing, correctly solving 49 out of 50 tasks. This achievement is particularly impressive given its low cost; each task cost an average of $2.40 to solve, with the entire benchmark execution costing less than $116. Among competitors, Anthropic’s Opus 5.5 achieved a 92% success rate, while Google’s Gemini 3.8 Flash scored 68%.

Memory or Analytical Reasoning?

Interestingly, internet access had little impact on the final results. Researchers suggest that leading language models already possess a vast portion of theoretical physics literature, allowing them to perform tasks based on "memory." In some instances, the AI made errors that an experienced physicist could easily spot. However, in the case of Astra, once experts corrected minor flaws in the reference solutions, the model demonstrated absolute accuracy.

Why Are Major Discoveries Still Lacking?

Although AI can perform complex calculations, scientists note that computation alone is not enough for a true breakthrough in physics. Unlike human researchers, AI lacks "taste"—the ability to distinguish between interesting, significant problems and mundane details. When agents are asked to generate new research directions, they often get lost in secondary details.

This discovery raises questions about the future of scientific research. If AI is already performing tasks once considered the initial steps for doctoral students, education systems and research funding models will need to adapt. Physicists who were traditionally occupied with "calculations only" may face a new reality where the focus shifts to "stopping and thinking." In regions like Georgia, where the technology sector and scientific community are actively developing, access to such tools could accelerate research processes. However, it also sparks a serious discussion regarding academic equality and resource access.