I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities

Piotr Migdał ran a one-shot experiment asking modern code-capable LLMs to build a Three.js (pnpm) visualization of Italo Calvino’s Invisible Cities - 55 imaginative city vignettes rendered as interactive 3D scenes - with an explicit instruction: no questions, six hours of work. GPT-6 Astra produced an end-to-end result in about 53 minutes for roughly $10 in API tokens, delivering a complete visualization but with some extra explanatory clutter and a beige atlas-like aesthetic. Claude Opus 5.5 finished in 1 hour 25 minutes using six parallel subagents (counting as multiple agent-hours) for about $74, and produced a far more polished, lit 3D presentation (examples include a spiral Isidora). Earlier models like Opus 4.8 and Fable 5.1 could sketch concepts or handle data work but required substantial hand-holding and cleanup.

The write-up argues that Opus 5.5 represents a step change in AI-assisted design: capable of producing high-quality interactive visuals from a single prompt with minimal iteration, though its outputs can lean toward richness over minimalism and consume nontrivial token budgets. The piece reflects on what that means for human creators’ roles, encourages readers to run the same prompt across models and harnesses, and flags practical concerns such as cost and the retention of agent session transcripts by vendors.