Physics benchmark CritPt v2 launches: GPT-6 Astra leads at 83.1% accuracy
OfirPress · x · 2026-10-22
After 4 months of expert review, the CritPt physics-reasoning benchmark ships v2 with 47 revised challenges: 5 problem-statement corrections, 19 revisions for underspecification, 16 grader fixes, and 7 refinements without issues found. GPT-6 Astra leads with 83.1% accuracy (average of 5 runs).
More from Models
- A Better, Fully Open Successor to BEIR for Retrieval Eval Is Coming — bclavie · 2026-10-22
- SmolVLM silently mangled images for ages as bf16 patch coordinates landed in wrong buckets — LysandreJik · 2026-10-22
- DeepSeek V4.1 Flash Finds CVEs at $2.40 Each, 27x Cheaper Than Opus 5 — teortaxesTex · 2026-10-22
- New psycholinguistics preprint dissects the weird emerging 'dialect' of frontier AI models — tallinzen · 2026-10-22
- OpenAI Codex usage limits reset today; team nudges users to set up Daybreak Blue — CurieuxExplorer · 2026-10-22
- Qwen4 and Kimi 3.1 reportedly spotted on Arena, unconfirmed — thatscoolbutno123 · 2026-10-22