GPT-6 Astra solves 49 of 50 unpublished physics problems at 98%, catching physicists' errors
141_1337 · reddit · 2026-10-15
A Reddit post claims GPT-6 Astra autonomously solved 49 of 50 unpublished theoretical physics research problems at 98% accuracy, under $3 per problem. Researchers initially marked several answers wrong, only to find the model had caught errors by world-class physicists. The attached chart shows realistic-track pass rates: Astra 98%, Opus 92%, Gemini 68%, DeepSeek 54%, GLM 42%. Third-party repost; original paper not directly linked.
More from Models
- First Proof batch 3: mathematicians grade GPT-6, Opus 5.5 and Kimi K3 on research-level proofs — littmath · 2026-10-15
- Google tops Signal65 PINNACLE agentic leaderboard: Gemini 4 Argon at $0.87 per correct task beats Sonnet and GPT — ryanshrout · 2026-10-15
- DeepSeek V4.1 Flash tops cost-efficiency at $0.134 per task, leaderboard shows — bindureddy · 2026-10-15
- Whatever Happened to Test-Time Scaling? Problems Needing pass@8192 Now Solve One-Shot — marc_lelarge · 2026-10-15
- cantrell on Gemini 4 Argon: heavy thinker, building long decode to push past 64K token limit — cantrell · 2026-10-15
- Opus 5.5 diagnoses fake OneDrive 'desktop takeover' in one prompt, maps 3 autostart paths — rehtorical · 2026-10-15