Anthropic's 'Most Dangerous' AI Builds Game in One Shot
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic released a model the source describes as 'supposedly too dangerous for the world to see' days before the June 12, 2026 article
- The author used a personal benchmark — building a game idea held for years in a single prompt — to test the new model
- The model ran a long reasoning session lasting 45 minutes and consuming over €20 worth of tokens
- Shepherd's Dog was produced as a single 2,319-line index.html file with zero dependencies, complete on the first try
- The author calls this the first time any AI model has successfully built this game in one go, linking a GitHub repo of failed attempts with earlier models as evidence
Why it matters: The model produced a complete 2,319-line game in one prompt after 45 minutes and over €20 in tokens — a result the author says no prior model had achieved on this benchmark. For developers testing AI coding limits, Anthropic's 'dangerous model' release doubles as a raw-capability demo, positioning the company as shipping maximum ability alongside its safety-restraint messaging.
Ask SkimNews
