Moonshot's Kimi K3 takes #1 on the Frontend Code Arena, beating Fable 5

Moonshot AI’s Kimi K3 took the top spot on Arena.ai’s Frontend Code Arena on July 16, the first Chinese model to get there. It scored 1,679 in blind human voting against Claude Fable 5’s 1,631 and GPT-5.6 Sol’s 1,618, and it won six of the seven frontend domains (Fable 5 held Gaming). The previous Kimi sat at #18 on the same board. And this is Arena, real people picking between outputs without knowing whose they are, not a self-graded benchmark table. The API runs $3 per million input tokens and $15 out; the Fable 5 credits it just beat cost $10 and $50. Moonshot says the weights go public July 27.

Before you plan a download: Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with a 1M-token context. That is GLM-5.2 territory, open in license and server-rack in practice, and no quant is dragging it onto a 24GB card. What an open frontier-grade coder means for this desk arrives downstream, in the distills that follow and in what a $3 frontier API does to everyone else’s pricing.

Then there is the reaction. David Sacks called the result “concerning” and blamed regulation: America is “tying itself in knots” while China runs permissionless. Worth remembering what one of those knots was. In June, Commerce put export controls on Fable 5 and Mythos 5 over a jailbreak Amazon researchers had reported, and Anthropic’s flagship went dark for 18 days while the company and the government argued over how serious it was. Sacks was the administration’s loudest voice for that move, accusing Anthropic of having “prioritized the continued offering of the consumer model over safety.” Five weeks later, the model that got benched is the model that got beat, and the same account is warning that slowing down American AI is how you lose the race. One of those knots still has his fingerprints on it.