One prompt, two models: Opus 5.5 and GPT-6 Astra build a 3D book site for $74 vs $10

Piotr Migdał of Quesma gave the same one-shot prompt to two coding agents: build a three.js visualisation of all 55 cities in Italo Calvino's novel Invisible Cities, using up to six hours. GPT-6 Astra in Codex finished end to end in 53 minutes at medium effort, for about $10 in API tokens. He says it worked but had design slop: extra labels, comments and visual noise. Claude Opus 5.5 in Claude Code took 1 hour 25 minutes with six subagents running in parallel (about seven agent-hours) and cost about $74. He called the result mesmerising, while noting it stopped well short of the six hours he asked for. This is one person's test on one task, not a benchmark. On Hacker News there was no consensus. Many found the site beautiful and impressive for a single prompt, and one commenter focused on value per dollar. Others pointed to telltale AI flaws such as bridges that go nowhere, reported slow performance on some browsers and phones, and several fans of the book felt it misses Calvino's point. One commenter added that Opus 5.5 ignored their own limit of 20 subagents and spawned 40. Sorted take: for design-heavy demos the pricier model may save cleanup time; for quick internal tools the cheaper run may be enough.