SI agent GPT-6-Astra solves 49 of 50 new theoretical physics problems, spending under $3 per problem
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An SI agent already solves theoretical physics problems at the level of specialists and for a few dollars, while deciding which questions are worth asking is still something only people can do.
According to The Decoder, 35 researchers from ICTS, Harvard, KU Leuven and Google DeepMind compiled TheoPhysEval — 50 unpublished research problems in theoretical physics with verifiable answers (calculation problems, no major open problems). The best result belongs to GPT-6-Astra in the Codex environment: 49 of 50 (98%), on average under $3 per problem and about $116 for the whole set. Claude Opus 5.5 in Claude Code scored 92%, Gemini 3.8 Flash 68%, and the open DeepSeek-V4.1-Flash and GLM 5.3 54% and 42%. Without internet access Astra and Opus each scored 92%: the authors take this as a sign that the models have largely memorized the literature. In a number of cases the SI "errors" turned out to be errors in the experts' reference solutions; after correction Astra's answers were right. The authors note that agents still lack "taste" — the ability to choose interesting problems — which is why there are no breakthroughs in physics like those in mathematics. The preprint was published on arXiv on October 9; The Decoder covered it on October 15.
Sources 2