CritPt v2: GPT-6 Astra hits 83% on research-level physics benchmark as frontier closes in
ArtificialAnlys · x · 2026-10-23
Artificial Analysis released CritPt v2 results — a research-level physics reasoning benchmark (70 composite challenges by 60+ physicists from Argonne/UIUC, with code execution in an offline Linux sandbox).
Key results:
- GPT-6 Astra (Max) leads at 83.1%, followed by GPT-6.1 Sol (Max) at 82.0% and Claude Opus 5.5 (Max) at 77.7%
- v2 scores aren't directly comparable to v1: GPT-6 Astra jumped from 32% to 83%, but the authors' testing shows most of the gain comes from the revised challenges (fixed errors and underspecified statements), with code-execution tool access adding a smaller boost
- With frontier models above 80%, CritPt v2 leaves less headroom than v1
Artificial Analysis will keep providing free grading API access to approved labs and researchers.
Related event: CritPt v2 Physics Benchmark Released; GPT-6 Astra Tops at 83%(4 posts)→
More from Models
- Models: useless or misleading for extrapolation, yet rock-solid on known-knowns — teodorio · 2026-10-23
- Diffusion LLM Mercury 3 beats Claude Sonnet 5.5 and GPT-6 Luna on simulated air traffic control — volokuleshov · 2026-10-23
- ThinkingCap squeezes 38.5% fewer thinking tokens out of Qwen3.8-Flash-Next at ~1pp accuracy cost — Thom_Wolf · 2026-10-23
- DeepSeek ships open-weight V4.1-Flash-1017, Terminal-Bench jumps 31.2 to 42.3 — petrusenko_max · 2026-10-23
- DeepSeek ships V4.1-Flash-1017: Terminal-Bench jumps 31.2→42.3, MIT-licensed — petrusenko_max · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23