The Kimi-K3 model achieved a score of 57 on the AAI Index, representing a notable improvement over its predecessor, K2.6. This new iteration also boasts nearly double the token efficiency compared to Opus 4.8, suggesting significant inference cost reductions.