GPT-6 Astra solved 49 of 50 physics research problems at $2.40 each in the new TheoPhysEval benchmark

#What happened

A team of 35 researchers — including contributors from ICTS, Harvard, KU Leuven and Google DeepMind — built TheoPhysEval, a new benchmark of 50 fresh theoretical-physics research problems with verifiable answers across 11 subfields. Commercial AI agents were set loose on the problems with no human hints.

GPT-6 Astra, running inside OpenAI's Codex coding environment, solved 49 of 50 problems (98%) at a cost of under $3 per problem — about $116 for the entire benchmark, roughly $2.40 per problem.

The runners-up: Anthropic's Opus 5.5 (via Claude Code) at 92%, Google's Gemini 3.8 Flash at 68%, DeepSeek-V4.1-Flash at 54% and GLM 5.3 at 42%.

Where the agent appeared to fail, some of the experts' own reference solutions turned out to contain errors; once corrected, Astra's answers checked out.

#Why it matters

The authors say leading models may already surpass the best human researchers on concrete physics calculations — a major milestone. Yet physics has not produced breakthroughs like math's Navier-Stokes result, because agents still lack "taste": the ability to ask the right research questions. The researchers expect autonomous top-level AI research sooner than most people assume.

Verified 2026-10-17