CritPt v2 contains 70 composite challenges, which models attempt using a code-execution tool in an offline Linux sandbox.
The underlying benchmark was written by more than 60 active physics researchers. According to Artificial Analysis, the benchmark authors and physics collaborators reviewed the v2 tasks to correct underspecified statements, errors, and typos. Artificial Analysis says it evaluates models independently and currently lists 11 models; its page reports GPT-6 Astra (Max) at 83.1%, GPT-6.1 Sol (Max) at 82.0%, and Claude Opus 5.5 (Max, Default Fallback) at 77.7%.
Artificial Analysis says it will provide a free grading API with the benchmark authors, available case by case to approved labs and researchers. The source does not establish that scores transfer to real scientific research, nor does it provide public API access for general users. Source: https://artificialanalysis.ai/evaluations/critpt-v2