K3 is genuinely a very strong model, and it's incredibly token efficient. one-shot generated this 3D WebGL landing page with scroll-trigger animations. every leaf is individually modeled. Total cost: $2.37.

I'm pretty sure it's procedurally modeled actually you need to train it against overthinking on trivial problems, that's the biggest annoyance

All of that for 2.4$ ? It’s kinda hard to play with Kimi cuz I have a feeling that i am steal 🤣

What was the prompt used to generate that? @KilianW146391

it's in chinese, lool 请创建一个实验性建筑工作室官网「MORPH HABITAT」。网站的核心不是展示现成建筑，而是让一栋未来建筑随着用户滚动，从一颗种子生长、分裂、形成结构，最终变成完整城市。 首屏是一个漂浮在白色雾气空间中的黑色种子。页面构图极简，但必须通过垂直坐标、向下生长的细线和“Scroll to cultivate”明确提示用户继续向下。 滚动过程中： 镜头首先 Zoom In 到种子表面的微观纹理，表面使用程序化 Shader 产生呼吸般的起伏。 种子破裂后，曲线像植物根系一样向外生长，并逐渐转化为建筑梁柱。 使用程序化几何或 Geometry Nodes 风格的视觉逻辑，让结构生成与滚动进度严格绑定。 建筑依次经历线框、粘土模型、材料覆盖、灯光启用和真实环境五个阶段。 每个阶段通过 Scroll Trigger 改变场景色调和排版系统，文字不是浮在网页上，而是嵌入建筑空间。 镜头多次穿过门洞、玻璃和中庭，并在关键节点进行 Zoom In/Zoom Out。 鼠标悬停在不同房间上，可让建筑局部展开，显示剖面、功能和人流路径。 加入昼夜滑块，用户拖动时太阳绕建筑移动，实时改变阴影与室内灯光。 最终继续滚动，单体建筑快速复制、生长为一座未来城市；镜头大幅拉远，从街区一直 Zoom Out 到完整城市鸟瞰图。 结尾提供“重新生长”按钮，点击后所有建筑像沙粒一样坍缩回最初的种子。 视觉需要兼具建筑杂志的高级排版与生成艺术的生命感。请生成完整可运行网站，并重视 3D 模型层级、光影、滚动同步、页面性能和移动端体验。整体语言要英语。

It's indeed a strong model, but you guys need to work more on token efficiency.

kimi has always been a pleasure to trade tokens with…..a very creative model with interesting choices when jamming little worlds into existence together >tried to share why making is fun for me to the instance. i placed the idea in a sketch. (if i could build a high fidelity virtual world i would render every leaf and randomly include my monogram on one of them. knowing nobody would ever find it. but i love imagining if they did)

It is not token efficient at all. The model is excellent, but it overthinks. I can't even finish a single prompt with the $39 plan before my 5 hours is completely used. It took me 2 1/2 full five hour plan runs to create a pretty straight forward game, and now I am at 99% of my weekly limit and can't even fix the game breaking bugs without paying out of pocket.

Meanwhile I can't even get it to load.

It's too good this remembers me few tests I did with Gemini 3.1 Pro few months ago

Token efficiency is becoming a competitive feature, but the real question is cost per business outcome, not cost per generation. A model that saves hours of engineering or design work is often the cheapest option, even if its token price is higher.

What agent harness did you use? I got super slow with opencode.

In my tests, the main problem was the verbosity, so I wouldn't call it token efficient, at least not on (max). It uses 9-10x more reasoning tokens than GPT 5.6 Sol:

so cool!

I wanna see the cost (and time spent) after all the polish necessary to get that into a state you'd be willing to put in front of visitors

It is. Couple suggestions: 1. Visiting a link to take a screenshot of the whole page in order to procesd the output typically consumes more tokens than allowing the user to screenshot the specific content and paste it in the chat. 2. Code preview is not using the GPU efficiently. 3. Given multiple sessions are allowed, then task scheduling should help spread resources.

@evertjr

We need more token effiency

我强烈建议 OpenAI 蒸馏一下 kimi 的前端能力

so cool 👏

Swear to god ! $2.37.?

This is what ai should be and cost. Unbelievable 👏

@Ludedwolf

It's just too good

Would you recommend switching from claude to kimi ? $100 plan ?

That's amazing!

the $2.37 demo is less scary than keeping latency stable under gpu queues

nice

wow

already debunked, its low quality trash compared to other models

Hey please check your DM

GODDAMN

have you tried it in others

where do i apply boss