Piotr Migdał gave Claude Opus 5.5 a single prompt — build a three.js visualization of all 55 Invisible Cities from Italo Calvino's novel — and a six-hour budget. The model used 85 minutes, deployed six parallel subagents totaling about seven agent-hours, burned $74 in API tokens, and produced something that stopped its creator mid-sentence. The interactive result is publicly available, and it's worth clicking before reading further. The experiment wasn't blind. Migdał ran the same prompt through GPT-6 Astra first, which completed in 53 minutes at medium effort for about $10 in tokens. That result was competent but sloppy — AI design slop, visual noise, Captain Obvious accessibility labels displayed verbatim, and a curious tendency to mimic Claude's aesthetic (beige backgrounds, zero-padded numbers). It was better than anything possible two years ago and worse than what a junior designer would ship. Opus 5.5's output landed in a different category. Migdał — who makes data visualizations and explorable explanations professionally — used the word 'mesmerized.' The piece traces a clear capability ladder: earlier projects with Claude Opus 4.8 on Proto-Indo-European etymology required heavy manual cleanup of AI slop and visual overlaps. Genetic Distance Map with Claude Fable 5.1 needed hand-holding on design. GPT-6 Astra was incrementally better at following design intent. Opus 5.5 was a step function. The catalyst was Ryan Sael's earlier demonstration: an interactive lens lab explaining camera focus, built by Opus 5.5 in one hour and 26 minutes for $25.66 in API costs, single-shot. Migdał saw it and didn't believe the consistency was real. So he chose a harder test — not optics, but literary architecture. Calvino's 55 cities are each an emotion rendered as urbanism. Translating that into interactive three.js is a design problem, not just a coding problem. The cost differential tells its own story. GPT-6 Astra: $10, 53 minutes, competent but noisy. Opus 5.5: $74, 85 minutes with six subagents, qualitatively different output. You're paying 7× more for something that arguably crosses a threshold from 'AI-assisted' to 'AI-designed.' Whether that threshold matters depends on whether you think design is selection or creation. Migdał's closing question — 'what is my place in creating interactive media?' — is the honest version of a question most creative technologists are asking privately. He's not performing anxiety; he builds these things for a living and just watched a model do it better than his previous collaborations with other models, in under two hours, from a single sentence. The piece is a capability report dressed as a blog post, and the capability being reported is taste. The implied trajectory is the real story. If Opus 4.8 required heavy cleanup, Fable 5.1 needed hand-holding, GPT-6 Astra was incrementally better, and Opus 5.5 is 'mesmerizing' — what does the next model do? Migdål says the ceiling is high and the next models will go higher. He's probably right, and that's either exciting or terrifying depending on which side of the prompt you sit on.
