GPT-6 Astra tops new physics benchmark, solving 49 of 50 problems
A team of 35 researchers from ICTS, Harvard, KU Leuven and Google DeepMind released TheoPhysEval, a benchmark of 50 unpublished theoretical physics problems with verifiable answers. OpenAI's GPT-6 Astra, running in Codex, solved 49 of 50 (98%) at under $3 per problem and about $116 for the full run; Anthropic's Claude Opus 5.5 in Claude Code scored 92%, Gemini 3.8 Flash 68%, and the open-weight DeepSeek-V4.1-Flash and GLM 5.3 reached 54% and 42%. The authors say web search made little difference and that agents still lack the "taste" to pick the right problems, so no comparable physics breakthrough has followed yet.
If the results hold up, frontier agents are already near expert level on verifiable routine research calculations, which could reshape how theoretical physicists train and who can afford to compete.