Hacker News front pageJun 13, 2026, 13:44 UTC+8

Building a full game in one shot with Anthropic's 'most dangerous' Claude Fable

Shepherd's Dog: A Game by the Most Dangerous AI Model

The author tested Anthropic's newly released Claude Fable on a game idea he'd held for years. After a 45-minute reasoning session costing over €20 in tokens, the model output a single 2,319-line index.html with zero dependencies—and the game just worked. He says this is the first time an AI pulled it off in one shot; earlier models all failed. The post doesn't name the exact model version or explain what 'dangerous' refers to in Anthropic's safety assessment.

Why it matters: The author tested Claude Fable on a game idea he'd had for years — after 45 minutes of reasoning and over €20 in tokens, the model produced a 2,319-line zero-dependency HTML game in one shot, where all previous models failed. It's a concrete, reproducible capability signal for...

Read the original ↗Export Markdown

Take

This one's worth opening because it gives a concrete upper bound for one-shot coding: not a function, but a full game idea the author held for years, turned into a playable index.html in a single run. 45 minutes of reasoning, over €20 in tokens, 2,319 lines with zero dependencies—those numbers put a price tag on "think hard, then write."

I'd discount two things. The post doesn't name the exact Claude Fable version, and it never explains what "dangerous" means in Anthropic's safety framing—the title is mostly playing off the launch narrative. Also, this is one sample; we don't know how complex the game logic really is or how maintainable the output would be.

The useful bit is the comparison: the author's GitHub repo logs earlier model failures on the same task, so this genuinely stumped previous systems. If you're trying to gauge whether long-reasoning models can replace the "stare at the problem for an afternoon, then code" workflow, this case is more telling than a benchmark score.

Source: Hacker News front page · koenvangilst.nl
Published: Jun 13, 2026, 13:44 UTC+8 · Jun 13, 2026