Kimi K3, and what we can still learn from the pelican benchmark

你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。

最容易做错的，是Kimi K3, and what we can still learn from the pelican benchmark；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：K3首个产品缺陷不是弱，是默认过度思考。

You see the launch news, almost scroll past, then stop because you do not want to miss the one detail that might change your next tool decision. My conservative take: K3's first product flaw is not weakness. It's default overthinking.

That matters because hype can hide cost. In Simon Willison's July 16, 2026 pelican test, the task was absurdly small: generate an SVG of a pelican riding a bicycle. K3 used 95 input tokens, 16,658 output tokens, and 13,241 of those were reasoning tokens, for about $0.25 on one tiny task.[S002]

The more important signal is how the product routes thinking, not just output quality. Moonshot says K3 launched with max thinking effort by default, while low and high controls are still coming later.[S001] So the first question is not whether K3 is smart. It is whether it thinks much harder than the task deserves.

A model update is worth your attention only if it changes your next decision, not because it lists more features. For beginners, that means testing the hidden token tax before you buy into the headline.

Boundary: this is based on one pelican SVG smoke test and the July 2026 launch setup, not a full workload comparison. If you are screening K3 for real work, start with a 30-second trivial task and check whether the thinking cost is doing real work for you. If you know someone about to judge K3 from the headline alone, share this with them before they spend time and budget on the wrong question.

真正该讨论的是：Kimi K3, and what we can still learn from the pelican benchmark