GPT-6-Astra Scores 98% on Theoretical Physics Benchmark

GPT-6-Astra solved 49 of 50 problems in TheoPhysEval, a benchmark for autonomous theoretical physics work built by 35 researchers from institutions including ICTS, Harvard, KU Leuven and Google DeepMind, according to their paper. The benchmark contains 50 unpublished problems with verifiable answers, each expected to take a specialist about one day or less without AI assistance.

The researchers gave agents access to literature searches and scripts for numerical and symbolic calculations, without human hints. GPT-6-Astra, running through OpenAI's Codex, scored 98% at an average cost below $3 per problem. Anthropic's Opus 5.5 scored 92%, while Google's Gemini 3.8 Flash scored 68%. DeepSeek-V4.1-Flash and GLM 5.3 scored 54% and 42%, respectively.

Without web search or access to background papers, GPT-6-Astra and Opus 5.5 each scored 92%. The researchers said the small change may indicate that the models retained substantial material from theoretical physics literature during training.

The problems cover 11 areas, including particle physics, cosmology, fluid flows and biological processes. They focus on concrete calculations using known methods rather than major open questions or the selection of new research directions.