Author Finds Claude Opus 5.5 'Mesmerizing' for One-Shot 'Invisible Cities' Visualization, Though It Used More Time and Cost Than GPT-6 Astra

One-Shot Design Test: Claude Opus 5.5 vs. GPT-6 Astra

A developer conducted a one-shot experiment to compare the design capabilities of Claude Opus 5.5 and GPT-6 Astra by prompting them to create an interactive three.js visualization of all 55 cities from Italo Calvino's Invisible Cities. The prompt instructed the models not to ask questions and to work for up to 6 hours until the result was a 'masterpiece'.

Performance Metrics and Results

Claude Opus 5.5 (via Claude Code):

• Time: 1 hour 25 minutes (utilizing 6 parallel subagents, totaling ~7 agent-hours).
• Cost: Approximately $74 in API tokens.
• Outcome: The author described the result as "mesmerizing" and a "wow!" moment, noting it as a significant jump in AI-assisted design quality. The model claimed to have used half the allotted time due to diminishing returns, though the author noted it finished early despite the instruction to use the full 6 hours.

GPT-6 Astra (via Codex):

• Time: 53 minutes at medium effort.
• Cost: Approximately $10 in API tokens.
• Outcome: The result was completed end-to-end but contained "AI design slop," such as unnecessary comments and visual noise. The author noted it seemed to mimic Claude's visual style (beige backgrounds, specific numbering) but lacked the aesthetic coherence of the Opus result.

Implications

The author concludes that while GPT-6 Astra is efficient, Claude Opus 5.5 represents a leap in interactive visualization quality from their perspective. However, the test highlights the high cost and time consumption of such high-quality outputs. The author questions their own role in creating interactive media given these capabilities, suggesting AI will become a primary tool for design-heavy workflows.