Kimi K2.6 Outshines Claude Design in Design Tasks

Model comparison

Kimi K2.6, a Chinese large language model, generated a more compelling movie‑grade scrolling indie‑game landing page than Anthropic’s Claude Design when given the same prompt.

Kimi’s API cost is reported to be about seven times lower than Claude Design.

Real‑world tests by @Aryan Rakib

First test: Prompted Kimi to design a high‑end “Velo City” cycling website. Kimi returned a complete landing page, UI mockup, site structure and product logic in a single response, which the tester said would normally require a designer, front‑end engineer and product manager.

Second test: Kimi generated a full‑stack site (front‑end and back‑end) without any user‑written code; the output was described as a deliverable product rather than a mockup.

Agent Swarm mode: Using Kimi’s multi‑agent clustering, the tester asked “Find 30 brick‑and‑mortar stores without websites and generate a landing page for each.” Kimi automatically split the task into parallel sub‑tasks and produced 30 customized designs simultaneously. Official data indicates the swarm can scale to 300 agents; the test used 30 agents.

Claude Design launch impact

Claude Design, powered by Opus 4.7, launched with a “one‑sentence‑to‑prototype” claim. On launch day Figma’s stock fell ~7 % and Adobe’s fell ~1.49 %; Anthropic CPO Mike Krieger resigned from Figma’s board three days before launch.

Three days later Kimi K2.6 was open‑sourced. Developers reported that Kimi’s design output surpassed Claude’s, attributing the advantage to Kimi’s strong general‑purpose foundation.

An independent Artificial Analysis benchmark ranks Kimi first among open‑source models worldwide, behind only the closed‑source “big three” (Claude, GPT, Gemini).

Tutorial: building a $10 000 website with Kimi

@Viktor Oddy demonstrated a workflow that required no code or design software, only conversational prompts.

Prompt example: “Create a modern brand site called Shamoni with a cat‑and‑flower theme, soft artistic style, including navigation, hero section, and intro module.” Kimi instantly generated a complete HTML page with branding, layout, typography and visual assets.

Iterative dialogue was used to adjust fonts, add card modules, tweak hover effects, implement a sticky navigation bar and scroll animations. Each instruction produced an immediate visual update, eliminating traditional design drafts, slicing and hand‑coded implementation.

Process summary: Describe requirement → Generate page → Iterate via language until satisfied . Compared with the classic pipeline (design → review → development → debug → revise), this approach removes at least three roles and streamlines communication.

Claude’s response

Anthropic formed an alliance with major creative‑software vendors (Blender, Autodesk, Adobe, Ableton, Splice) and released a suite of connectors that embed Claude directly into those tools via Blender’s Python API, funded through the Blender development fund.

Claude pursues a “two‑pronged” strategy: a native AI design tool (Claude Design) and AI interfaces for existing design applications.

Kimi K2.6 remains a single open‑source model that, with a prompt and a cheaper API bill, can take an idea from concept to a live site without requiring designers or developers to switch tools.

Key quantitative points

Kimi’s price ≈ 1/7 of Claude Design.

Agent Swarm can scale to 300 agents; test used 30 agents.

Artificial Analysis benchmark: Kimi #1 among open‑source models, only behind Claude, GPT, Gemini.

Reference URLs: https://x.com/socialwithaayan/status/2049094873186435260, https://x.com/HeyZoyaKhan/status/2049197606719623372, https://x.com/viktoroddy/status/2049069846953103413, https://x.com/tec_aryan/status/2049170121349062941.

Code example

往
期
推
荐
1、
什么操作？改一行前端代码，竟然惊动了四个后端团队！！
2、
面试官问我：“AI 写代码比你快 100 倍，你的价值在哪？”
3、
原来有这么多人把 OpenAI/Claude 的 API Key 直接推上 GitHub？
4、
从卡顿到丝滑：Spring Boot 接入 Redis 缓存的正确打开方式
5、
原来有这么多人把 OpenAI/Claude 的 API Key 直接推上 GitHub？
6、
面试官：“为什么敏感词过滤不用暴力匹配？” 我：“用暴力匹配的同事性能已经挂了”
点
分
享
点
收
点
点
赞
点在看