Anthropic released an AI model described as too dangerous for public use, which successfully created a complete game called Shepherd’s Dog in one session. The game, consisting of a single 2,319-line index.html file with no dependencies, was developed over 45 minutes using more than €20 worth of tokens, according to Koen van Gilst’s personal benchmark test on June 12, 2026.
Van Gilst tested the model by asking it to build a game idea he had envisioned for years. The AI engaged in an extended reasoning process before delivering the final product, which matched his expectations exactly. He shared a video demonstration and made the game available to play online, highlighting the AI’s ability to produce complex, functional code in one attempt.
This achievement marks a notable step in AI capabilities, demonstrating how advanced models can generate fully operational software without iterative human input. The game’s creation in a single run contrasts with previous models that required multiple attempts, showcasing progress in AI’s coding proficiency. Such developments could influence future AI applications in software development and creative industries.
The Shepherd’s Dog game can be played at vnglst.github.io/when-ai-fails/shepards-dog/claude-fable-5/index.html, providing a tangible example of the AI’s output. Van Gilst’s documentation of earlier attempts is available on GitHub, offering insight into the model’s evolution and performance.