GPT-6 Astra Solves 49 of 50 Physics Problems on New Benchmark

A team of 35 researchers, including scientists from ICTS, the Tata Institute of Fundamental Research, the University of Pennsylvania and Google DeepMind, has published a benchmark called TheoPhysEval that tests how autonomously AI agents solve theoretical physics research problems using web search and code execution tools, according to aitimes.com. On 50 unpublished theoretical physics problems, GPT-6 Astra solved 49 without human hints.

OpenAI's model ran at a total API cost of $116.28, an average of about $2.33 per problem. Opus 5.5 solved 46 of the 50 problems, Gemini 3.8 Flash 34, DeepSeek-V4.1 Flash 27 and GLM-5.3 21.

Blocking internet access changed the picture only slightly: GPT-6 Astra and Opus 5.5 each solved 46 problems under those conditions.

One problem defeated every model

GPT-6 Astra failed a single problem, finding the differential equation that generates a given curve — and all the other models failed it as well. During grading, answers the model had initially been marked wrong on turned out to be correct after experts rechecked and fixed subtle errors in their own reference solutions, lifting its final accuracy to 98%.

Co-author Subhraj Raju of ICTS wrote on his personal blog on the 14th about how the evaluation began, and said he could not sleep for days after seeing the results. Economist Alex Tabarrok also commented on the results on his blog Marginal Revolution. About six months before the article, Raju gave an unpublished problem to an AI, which answered incorrectly at first but solved it after several hints.