Agreed. The benchmark closest to my experience is FrontierMath Tier 4. Fable and Sol (90%) are very far ahead of Kimi K3 (not even 40%). Kimi is trained heavily to basic agentic tasks, like all the other open models right now.