GPT-6 Astra solves 49 of 50 research-level physics problems, catches experts' own errors
MoonL88537 · x · 2026-10-24
TheoPhysEval, a new benchmark from 30+ theoretical physicists on arXiv (Oct 9), gave five agents 50 unpublished research-level physics problems. GPT-6 Astra in OpenAI's Codex harness answered 49/50 correctly at under $3 per problem. Initial grading put it in the high 80s until the model insisted the expert answer keys were wrong — physicists rechecked and found subtle errors in their own calculations, and after fixes Astra scored 49/50. 'The AI ended up testing the humans,' as Alex Tabarrok put it.
More from Models
- Wave of Chinese Model Releases Imminent: GLM-5.5 and Qwen 4 Next Up — haider1 · 2026-10-24
- Kimi K3.1 Pro jumps 27.6 points to enter BuildingBench top six, matching DeepSeek V4.1 Flash — Lianhuiq · 2026-10-24
- Kimi's $100 tier gives less usage than a $20 Claude sub, yet thrives in China — Linahuaa · 2026-10-24
- humans& ships collaborator model Ampersand, social product Boba, and open-sources 10% FTL decode speedup — ericzelikman · 2026-10-24
- Nat Eliason: Blind test 'OVERWHELMINGLY' picks Argon's writing, wants it to rewrite everything — nateliason · 2026-10-24
- Video: Kimi K3.1 Pro benchmarks claim wins over Opus and GPT flagship, but is it benchmaxxed? — TheAIGRID · 2026-10-24