Star PandaCode on GitHub: https://github.com/aicodeking/pandacode
Join this channel to get access to perks:
https://www.youtube.com/channel/@aicodeking/join
In this video, I'll be testing Moonshot's new Kimi K3.1 Pro on all six KingBench 4 tasks, and I'll compare it with Opus 5.5, GPT-6 Astra, Haiku 5.5, DeepSeek V4.1 Flash and the rest of my chart. I ran the whole benchmark through PandaCode, my fully open-source app for coding agents, which is tuned for Kimi Code, with six K3.1 workers running in parallel and a separate evaluator testing every project. We'll go through the emulator, the Docker alternative, the room designer, the keyboard app, Pokémon Red and the Fitbit app one by one, and then look at the final score, what the run actually cost and what the open weights mean. At the end, I'll tell you whether K3.1 Pro is worth using, because the launch blog's "almost Opus" story does not hold up in real projects.
--
Key Takeaways:
🚀 Kimi K3.1 Pro is Moonshot's post-trained K3 with a 1M context, a 2–3× faster Highspeed version, and open weights promised by November 4th.
📊 The launch blog claims 65.7% on Terminal-Bench 4.0, right next to Opus 5.5 at 66.4% and well ahead of GPT-6 Astra at 58.2%.
💻 I ran all six KingBench 4 tasks through PandaCode, which is fully open source and has a dedicated Kimi Code integration.
🧠 K3.1 scores 5.37 out of 10, sixth out of eleven, behind Opus, Astra, Sol, Argon and even Haiku 5.5.
✅ It is honest about what it didn't build, it blocked all 20 malicious images in the Docker task, and it speaks the real VIA protocol in the keyboard app.
⚠️ It built only the Game Boy part of the emulator, its Docker alternative has no build or networking, and Pokémon has broken levelling and a save-wiping title screen.
💸 The whole run cost about $31.88, compared with $7.40 for Haiku 5.5 and under $2 for DeepSeek V4.1 Flash.
🔍 Most of that bill is 110 million cached tokens that it re-reads on every step, not the code it writes.
🎯 Once the weights land, it will be the best-scoring open-weight model on my chart, ahead of DeepSeek Flash and GLM.
👍 Verdict: a good open model, not a frontier one, and on the API Haiku or DeepSeek Flash are the smarter picks for now.
Moonshot 的發表文說 Kimi K3.1 Pro 幾乎跟 Opus 5.5 平起平坐，我本來還想說開源模型終於追上來了，結果 AICodeKing 跑完六項測試只給 5.37 分，排第六，連 Haiku 5.5 都輸。這支影片就是把這個「幾乎 Opus」的故事拆開來看，一項一項對照它到底做到什麼、沒做到什麼。
