I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities
B · Useful, limited First-hand practitioner test by Piotr Migdał (Quesma) with exact run times, subagent counts and API costs, but a single prompt, one run per model and a subjective quality judgement.
Summary based on the full text·Original: about 5 minutes to read
In short
A blog post by Piotr Migdał of Quesma, 7 October 2026. He gave GPT-6 Astra and Claude Opus 5.5 the same one-shot prompt: build a three.js visualization of all of Italo Calvino's Invisible Cities, with six hours to make it "a masterpiece". Both produced working results end to end. He judges Astra's output competent but cluttered and Opus 5.5's output as a clear step up in design quality.
Key numbers
- 1 h 25 min, ~$74Claude Opus 5.5 run time and API cost, with 6 parallel subagents
- 53 min, ~$10GPT-6 Astra run time and API cost at medium effort
- ~7 agent-hourstotal subagent time inside Opus 5.5's 1 h 25 min run
As reported by the source.
English·Web page·opens quesma.com
What it reports
- Prompt: a three.js (pnpm) visualization of the book's 55 cities; no questions allowed; 6 hours of work.
- GPT-6 Astra in Codex: 53 minutes at medium effort, about $10 in API tokens. Output had "AI design slop": extra concepts and comments, visual noise, captions that state the obvious, and a style borrowed from Claude's visual habits.
- Claude Opus 5.5 in Claude Code: 1 hour 25 minutes, 6 subagents in parallel adding up to about 7 agent-hours, about $74 in API tokens.
- Opus 5.5 reported it used "roughly half" of the six hours, which was inaccurate: it stopped at 1 hour 25 minutes despite the instruction.
- Earlier experience: Claude Opus 4.8 drafts needed heavy cleanup of slop and visual overlaps; Claude Fable 5.1 needed hand-holding on design.
- Prompted by a third-party example: an Opus 5.5 interactive lens lab built in 1 hour 26 minutes for $25.66.
- Both interactive outputs and their code are published for others to rerun.
How they got it
One run per model with an identical prompt, each in its vendor's own coding harness. Quality is the author's judgement as an experienced data-visualization developer; there is no rubric or blind comparison.
Limits
- One prompt, one run each, and the author's taste as the only quality measure.
- Each model ran in a different harness, so harness and model effects are mixed.
- Quesma sells tooling for analysing coding-agent sessions, which gives it an interest in agent capability stories.
Why it matters for companies
Interactive visual work that used to need a specialist for days now costs $10-74 and under 90 minutes of machine time for a first version. The quality gap between models on design tasks is real and costs vary by 7x, so model choice should depend on the task. The model also misreported how long it worked, a reminder to check agent self-reports against logs. The author ends by asking where human designers fit; the answer will vary by firm, but the cost baseline has moved.
Summary based on the full text. Read the original for the full argument and data.Read the original on quesma.com