Anthropic recently released a model deemed too risky for public use, yet it successfully crafted a game called *Shepherd's Dog* in a single attempt. The AI spent over €20 in tokens and 45 minutes to generate a 2,319-line index.html file with no dependencies. This achievement marks a rare instance where an AI delivered a complex project exactly as envisioned, bypassing traditional development hurdles. The game is playable at vnglst.github.io/when-ai-fails/shepards-dog/claude-fable-5/index.html, demonstrating both technical prowess and the model’s unpredictable capabilities.

The experiment stemmed from a personal challenge: testing whether the AI could execute a long-held game concept without iterative feedback. Earlier models failed, but the latest version executed the task flawlessly. The €20 cost highlights the financial weight of token usage, a critical factor for developers. This success raises questions about AI’s role in creative coding and whether such efficiency could disrupt traditional workflows. While the model’s danger remains debated, its ability to generate functional code in one go is undeniably significant.

The *Shepherd’s Dog* game itself is a testament to the model’s unexpected potential. It operates without external libraries, relying solely on raw HTML, which simplifies deployment. For developers, this could mean faster prototyping or reduced reliance on human coders. However, the experiment also underscores risks—AI-generated code may lack nuance or security. The fact that the model handled a complex project in a single shot challenges assumptions about AI’s limitations. This isn’t just a technical feat; it’s a preview of how AI might reshape software creation, forcing the industry to rethink safety, cost, and control.