Dutch developer Koen van Gilst handed his long imagined 'Shepherd's Dog' to Anthropic's newly released model, which the company had previously described as too dangerous to ship, in a single prompt.
For a decade, Koen van Gilst had a game in his head. "Shepherd's Dog" imagined a player tending a flock through weather and wolves, in a single hand-built HTML file with no external dependencies. He had tried to get AI to build it before, with earlier models, and the output always broke or needed too many repairs to feel like his. Then Anthropic released a new model, one the company had previously described as too dangerous to release, and van Gilst decided to make it his personal benchmark: one prompt, no iterations, no second chances.
The result, according to van Gilst's write-up of the experiment, was a single 2,319-line index.html with no dependencies, produced in roughly 45 minutes of generation time and at a token cost of more than €20. The model, which the source identifies only as "Claude Fable 5", a name that does not match Anthropic's publicly known model lineup and may be a codename, an internal handle, or a placeholder, reportedly returned the line "Shepherd's Dog is complete" when it finished.
Anthropic's own positioning of the model as too dangerous to release is vendor language, not an independent capability assessment. What makes the test useful, and worth reporting, is the setup rather than the slogan. Van Gilst had a long-held personal benchmark, a public record of prior failures with earlier models, and a willingness to pay the actual cost of generation. That gives the result texture in a way most "AI can now code a game" announcements do not.
The honest cost matters. Forty-five minutes is a long time to wait for one prompt. More than €20 is a non-trivial price for a personal experiment. Both are part of the constructive truth of where the one-shot bar sits right now, not footnotes. Van Gilst's earlier attempts, archived in the same GitHub repository he links from the write-up, show why a single working result is a real shift. Prior models could not produce the game in one go.
Van Gilst describes the output as "really fun and exactly how I imagined it," and frames it as the first time any model has produced the game for him in a single attempt. That is a personal verdict from a single developer, not a benchmark. No third party has reviewed the code, run an independent playtest, or measured the artifact against comparable one-shot attempts by other models. Anyone deciding whether to call this a capability milestone should treat the artifact and the cost, not the "too dangerous" framing, as the evidence.
The model name itself is the loose end. "Claude Fable 5" does not appear in Anthropic's public model catalog, and the company's actual safety and release framing for the underlying system is unverified in the source material. If the model's real name and safety positioning are softer than the slogan suggests, the lead of the story still holds. A long-held personal project crossed the one-shot bar for the first time, at a real cost in time and money. The headline-worthy part is the creator's receipts, not the company's marketing.