Harvard and others release TheoPhysEval: GPT-6 Astra scores 98% after experts revise answers
哈佛谷歌惊人研究:GPT-6 Astra已解决98%世界物理难题!
The TheoPhysEval evaluation, involving institutions such as Harvard and Google DeepMind, shows that GPT-6 Astra scored about 88% in the initial assessment on 50 unpublished theoretical physics problems, rising to 98% after experts reviewed and corrected the reference answers, solving 49 problems. The problems were drawn from experts' unpublished research notes, covering quantum field theory, holographic theory, and more, and included two test tracks, one with internet access and one without; offline, both Astra and Claude Opus 5.5 scored 92%. The report states that the average API cost per problem was under $3. A concurrent MIT team study shows that frontier AI's problem-solving rate in theoretical science rose from 2% in mid-2024 to 35% in mid-2026.