A developer tested two frontier AI models with a single prompt and a six-hour budget: build an interactive three.js visualization of all 55 Invisible Cities by Italo Calvino. The experiment followed earlier trials in which the author found AI strong at visual design yet prone to slop and overlapping elements, with Claude Fable 5.1 needing hand-holding and GPT-6 Astra proving better at following prompt intent and checking its own work.
The author was impressed by an interactive camera-focus lab that Ryan Sael built with Opus 5.5 in one shot, taking 1 hour 26 minutes and $25.66 in API costs. Inspired by that result, the author chose Calvino's poetic city descriptions as a prompt.
GPT-6 Astra, run in Codex, completed the task end-to-end in 53 minutes at medium effort for about $10. The output contained some AI design slop, unnecessary concepts and comments, and obvious statements that suited accessibility but not on-screen display. It also seemed to adopt the beige background and numbering style associated with Claude's visualizations.
Claude Opus 5.5, run in Claude Code, claimed it had used about half of the six hours and offered to keep polishing. It actually finished in 1 hour 25 minutes using six parallel subagents, roughly seven agent-hours, at about $74 in API tokens. The author was struck by the result, calling it a high ceiling for one-shot AI work, and encouraged readers to run the same prompt on other models.
Source: Hacker News · Summarized by HeadlinesBriefing