Shepherd's Dog: A Game by the Most Dangerous AI Model
摘要
作者测试了 Anthropic 的 Claude Fable 模型，尝试让其在单次提示下实现构思多年的游戏《Shepherd's Dog》。该模型在经过 45 分钟的推理并消耗超过 20 欧元的 Token 后，交付了一个包含 2319 行代码、零依赖的 HTML 文件。作者表示这是 AI 首次能一次性精准完成此类复杂开发任务，且结果完全符合预期。
荐读理由
你可以参考文中对长推理模型在单次生成任务中的量化表现（45分钟推理、2300行代码、20欧成本），预判 AI 编程从对话式迭代向长逻辑推理、单次成型模式演进的技术信号，从而校准你对未来 AI 工程交付边界的判断。
原文
Shepherd's Dog: A Game by the World's Most Dangerous AI
June 12, 2026
1 min read
A few days ago, Anthropic released the model that was supposedly too dangerous for the world to see. I wanted to give it a try using my personal benchmark: can it build a game idea I've had for years, in a single shot?
The model started with a very long reasoning session, and after 45 minutes and more than €20 worth of tokens (!), it happily reported that "Shepherd's Dog is complete — a single 2,319 line index.html with zero dependencies."
You can watch the video below or give the game a try It's really fun and exactly how I imagined it. For me this is the first time an AI model is able to create this for me in one go:
Play the game here: vnglst.github.io/when-ai-fails/shepards-dog/claude-fable-5/index.html
Other attempts with earlier models: github.com/vnglst/when-ai-fails/blob/main/shepards-dog/README.md
这条对你有帮助吗？