Kimi K3 Max

Kimi K3 Max is not the name of a model

The row you see on the Arena leaderboard is called kimi-k3-max, but Moonshot's model catalogue lists no model by that name. It lists kimi-k3. The max suffix is a setting, the amount of effort the model spends reasoning, not a separate product with a separate price.

The same pattern repeats on GLM-5.2 and on DeepSeek, and most comparison lists treat each row as its own model. We kept the slug people actually search for, and attached the price and the context window to the base model, because the source that publishes those two numbers is talking about the base model.

The price, and a second rate that only applies to repeated prompts

Three dollars per million input tokens and fifteen per million output. Against Opus 5 at $5 and $25, that is forty percent less. Moonshot publishes a second input rate too: when the prompt is already cached, input bills at $0.30, a tenth of the standard rate.

We did not put that $0.30 in the price column, for a plain reason. No real workload is one hundred percent cache. If your architecture has a long fixed prefix repeated a thousand times a day, the cached rate is real for you and you should work it out yourself. If every request is fresh, three dollars is what you pay.

Why it comes third rather than second in our table

Kimi trails Fable 5 by two points in the coding table, and all two of them come out of one column: tooling maturity. Fable 5 takes full marks and the whole fifteen points of weight there; Kimi takes zero, because Moonshot's platform documentation lists neither an official command line tool nor an editor extension. On the leaderboard itself, Kimi is ahead of Fable 5.

So if your team writes its own tooling layer and does not need a ready-made agent SDK, those fifteen points mean nothing to you, and Kimi effectively sits higher than our table shows. The table cannot know that about you. You can.

The column Kimi wins outright, and why we distrust it

Kimi's context window is 1,048,576 tokens. That is two to the twentieth, not a rounded marketing million. Every other candidate with a published window sits at exactly one million, so a difference of under five percent hands Kimi all ten points of the context column and everyone else zero.

Min-max normalisation does that, and we do not bend it, because the moment one column gets adjusted by hand the rest of the table stops meaning anything. But a reader should know those ten points are not ten points of real advantage. In practice the two windows are the same size.