GPT-6-Astra solves 98% of 50 unpublished theoretical physics problems at under $3 each
RexDouglass · x · 2026-10-15
Theoretical physicist Suvrat Raju and colleagues built a benchmark of 50 unpublished research-level theoretical physics problems with verifiable answers—each typically takes a world expert about a day. With realistic tooling (literature search, numerical/symbolic scripting), the best agent, GPT-6-Astra harnessed in Codex, hits 98% accuracy at under $3 per problem. Accuracy barely drops without web access, suggesting substantial literature memorization. The paper argues frontier agents can now handle concrete calculations across theoretical physics, and discusses why agent-led breakthroughs remain rare.
More from Models
- Logit lens shows Qwen3-1.7B's refusal to be shut down depends on how many layers you train — Sauers_ · 2026-10-15
- Reflection AI's 501B-parameter Beam-1.0 model spotted in early access — jasonkneen · 2026-10-15
- Jev was a multimodal image model all along, just disabled via a 'Danger Zone' header — triviatroy · 2026-10-15
- Claire Vo benches Meta's new Muse Wonder: strong on agents, weak on design — alexandr_wang · 2026-10-15
- Lab Beef Over 'Show Your Trajectory': Skeptics Call Argon Benchmark Claim Into Question — teortaxesTex · 2026-10-15
- Banned hours after renewing: Claude Max ban lottery strikes again — yangyi · 2026-10-15