Artificial Analysis determined that Kimi K3.1 consumes 14% more tokens per task than Kimi K3