Artificial Analysis launches and regrades CritPt v2 physics benchmark
Article timeline
The articles in this story. Times are ET.
Artificial AnalysisOfficialArtificial Analysis finalizes CritPt v2 grader, releases Intelligence Index v5.0.1AIArtificial Analysis says it worked with the CritPt authors to finalize the CritPt v2 grading server, which will be available through its API tomorrow. The authors' robustness fixes were incorporated as v5.0.1 of the Artificial Analysis Intelligence Index. After regrading all models, most scores rose slightly, but frontier model rankings are unchanged and no score moved by more than 1 point.
- Artificial AnalysisOfficialPickArtificial Analysis launches CritPt v2, a revised physics benchmark run with code execution
AIArtificial Analysis has launched CritPt v2, a revised research-level physics benchmark after its authors reviewed all 70 challenges and corrected errors and typos. Models now run agentically, writing and executing code in an offline sandbox. GPT-6 Astra (max) scores 83% and GPT-6.1 Sol (max) 82%, while GPT-5 (high), released in August 2025, scores 29%.
- Artificial AnalysisOfficialPickArtificial Analysis launches CritPt v2, a revised physics benchmark with code execution
AIArtificial Analysis says it has launched CritPt v2, a revised version of the research-level physics benchmark. The team reviewed all 70 challenges, corrected errors and typos, and now lets models write and run code in an offline sandbox. GPT-6 Astra (max) scores 83%, and GPT-5 (high), released in August 2025, scores 29%.
