Solving a task once does not mean a model will answer it consistently, and consistency is a key part of the advantage for frontier models on VIALS
Kimi K3.1 Pro answers 119 of 161 tasks correctly at least once across three attempts, compared with 118 of 161 tasks for GPT-6 Astra. Requiring all three attempts to succeed changes the picture: Kimi passes 73 tasks (45%), while Astra passes 101 (63%)
