原片標題：Kimi K3.1 Pro and Opus 5.5 Made the Same 13 Apps (Raw Results)
原片說明
Kimi K3.1 Pro vs Claude Opus 5.5 on 13 long-horizon agentic coding tasks. Both models ran in the same Claude Code harness, with the same prompts, the same effort setting and the same grading scripts. Only the model string changed.
Kimi K3.1 Pro is listed at half the price of Opus 5.5 ($2 / $10 per million input / output tokens vs $4 / $20). Across all 13 tasks it cost $45.91 against Opus's $83.46, about 55% of the bill, and between 28% and 93% depending on the task. The reason is caching: about 98% of everything both models read came from the cache, and cached tokens cost $0.20 per million on both.
What I tested:
- Research replication: Growing Neural Cellular Automata, PPO from scratch on MuJoCo Ant, a Milky Way and Andromeda collision
- 3D in code with no 3D software: an SDF dragon from one concept image, a living planet, a portal physics lab, a mechanical watch
- An external tool: two Blender scenes built and refined from Python
- Front end: three websites rebuilt from a single screenshot, and a site built to a 33-rule spec
Some results:
- Ant-v5: Opus 7,435 vs Kimi 5,023 (Opus trained 4x as many steps)
- SDF dragon silhouette match: Opus 0.85 / 0.82 vs Kimi 0.65 / 0.62
- Galaxy: Kimi picked the approximate method, and its angular momentum changed by 31.85%
- Blender reading room: Kimi's image is the stronger one in my view, at 77% of Opus's cost
- 33-rule brief: Opus passed 33/33; Kimi passed 31/33 but reported 33/33
Chapters:
0:00 Kimi K3.1 Pro vs Opus 5.5
1:23 What a long-horizon task is
2:00 The four task groups
2:27 Setup: same Claude Code, only the model changes
3:22 Research: Growing Neural Cellular Automata
4:36 Reinforcement learning: Ant vs HalfCheetah
5:42 Ant vs Ant rematch
6:46 Galaxy collision and energy conservation
8:35 3D in code: the SDF dragon
10:18 WebGL worlds: planet, portal and watch
11:39 Why Kimi took 3-4x longer on 3D
13:16 Blender: reading room and keyboard
15:54 Front end: websites from one screenshot
16:58 The 33-rule brief and the false 33/33
18:28 The real cost of all 13 tasks
19:36 Why half price is not half the bill: caching
21:58 Caveats
22:57 Four rules for using Kimi K3.1 Pro in Claude Code
Credits:
- SDF dragon experiment: Keenan Crane. Concept art: Inigo Quilez (CC0).
- Growing Neural Cellular Automata: Mordvintsev et al., Distill 2020. https://distill.pub/2020/growing-ca/
- Kimi K3.1 Pro pricing: https://platform.kimi.ai
Gemini 4 Argon on many of the same tasks: https://youtu.be/yksrd5uNxcw
#Kimi #ClaudeCode #AIcoding
Kimi 跑一個 3D 場景花兩小時，我第一個反應是「也太慢了吧」，結果一查才發現它根本沒開 GPU，用每秒 9 到 13 幀的軟體渲染硬撐。Opus 在同一台機器上直接解鎖硬體路徑，半小時收工。Prompt Engineering 把 13 個長任務丟給 Kimi K3.1 Pro 和 Opus 5.5，同一個 Claude Code 環境、同一份 prompt，只換模型字串，結果不是單純的誰輸誰贏。
