Below is Kimi K3's version of the Anchor 3D task I had Fable try a few days ago. We replicated the setup to be exactly the same as before (Anchor2 checkpoint), and then I had Fable run the process itself. Seems like Kimi did the job correctly, perhaps even slightly more tastefully than Fable. The version you see in the homepage here is actually Fable's port of Kimi's solution to the website's engine, but Kimi's original is here https://a327ex.com/media/shared/games/kimi-k3-playground/index.html. Fable actually had to do some work to capture the feel of Kimi's version, which I thought to be slightly superior to Fable's, but these are small details, both models did the job correctly.

More pressingly though is the fact that this task alone cost me 50% of Kimi's weekly usage!!!! It also took a long time. Fable maybe took 20 minutes, I don't remember, but this one easily took like an hour for Kimi. This was because I used max reasoning effort by default instead of setting it lower. At the same time, for a fair comparison, there's no equivalent to Fable's "extra" for Kimi, which is what I used, so yea, in the end everything cost more and took longer. It's not really a perfectly fair comparison but it's as close as I could get.

I will use the rest of my weekly Kimi usage on actual projects instead of trying to faithfully do work Fable has already done to see which one does it better. For my use cases, I think it's clear both models are very capable. I'm mostly curious about whether Kimi is a better gamedev companion than Claude now. Given that it's slower I doubt that will be the case, but maybe it has significantly better taste such that its creativity outshines its slowness, we'll see.