Artificial AnalysisT1·58 pts
- Private dataset share rose from 25% to 45%, adding the 1,500-document AA-Paperclip benchmark at 30% weight
- The new Decision Score rewards both correctness and calibrated confidence, scoring above 50 for right answers and below 50 for wrong ones
- Mammoth 27B is within 4 points of GPT-6 Luna while running 118x faster