GPT-6 Astra solved 49 of 50 unpublished theoretical-physics problems in TheoPhysEval, a benchmark assembled by 35 researchers from institutions including Harvard, KU Leuven, Google DeepMind and the International Centre for Theoretical Sciences. Its full run cost just under $116, or less than $3 per problem.
The collection spans 11 subfields and uses verifiable calculations that a specialist could normally complete within about a day. Contributors often changed values from recent work so the answer could not simply be copied. The benchmark intentionally excludes grand open questions and instead measures routine “turn-the-crank” research tasks.
Astra scored 98% with web search and 92% without it. Anthropic’s Opus 5.5 reached 92%, while Gemini 3.8 Flash, DeepSeek-V4.1-Flash and GLM 5.3 scored 68%, 54% and 42%. Reviewers also found several errors in the original answer key after agents produced conflicting results.
The strongest score does not demonstrate independent scientific discovery. Researchers still need to formulate theories, select worthwhile questions and recognize when a harder problem should be reframed. The benchmark’s authors describe that judgment as scientific “taste,” which current agents lack. The result is stronger evidence for automating specialized calculations than for replacing the broader research process.