DeepMind-backed benchmark: best AI agent solves 49/50 unpublished research-level physics problems
rohanpaul_ai · x · 2026-10-25
Google DeepMind and many top universities released TheoPhysEval, a suite of 50 unpublished research-level theoretical physics problems (each roughly a day's work for a world expert) to test off-the-shelf AI agents autonomously.
- The best agent, GPT-6-Astra in Codex, solved 49/50 with access to web search and scripting, at under $3 per problem
- Both GPT-6-Astra and Opus 5.5 in Claude Code solved 46/50 without web search
- Accuracy barely dropped without search, suggesting the models have memorized much of the physics literature
More from Models
- Dev: DeepSeek is very hard to beat on easy tasks, 10x cheaper than Sonnet — bindureddy · 2026-10-25
- OpenAI's rumored 6.1 Terra points to demand for smaller, cheaper models — haider1 · 2026-10-25
- SRE Bench: GPT-6 scores 57% at binary reverse engineering, DeepSeek models under 2% — teortaxesTex · 2026-10-25
- MOSS-Image: 5B open image model hits 1.3M downloads in four days — neurocartel · 2026-10-25
- Why DeepSeek's Flash models dazzle but Pro versions disappoint: a training-signal theory — teortaxesTex · 2026-10-25
- Pro user: 62,500 Codex credits ($2,500) lock to smallest models the day your plan ends — lanqo88 · 2026-10-25