I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities
Visualizing Invisible Cities: A Six-Hour Experiment with Opus 5.5
By Piotr Migdał | October 7, 2026
For someone dedicated to the craft of explorable explanations and data visualization, the rise of Large Language Models (LLMs) has been a paradoxical experience—simultaneously a gift and a source of irritation.
While even basic local models can now whip up a decent storefront landing page, the "last mile" of design remains tricky. In my previous work, such as the Tree of ‘tree’ (built with Claude Opus 4.8) or a Genetic Distance Map (via Claude Fable 5.1), the initial outputs were impressive but required significant manual labor. I spent hours scrubbing AI slop and correcting visual overlaps to ensure the design actually felt "good."
The Catalyst
My curiosity was piqued by a post from Ryan Sael on X, who showcased an interactive lens lab designed to explain camera focus.
"I asked Opus 5.5 to explain camera focus by building an interactive lens lab. Here’s what it came up with after 1 hour 26 minutes in one shot, $25.66 API cost."
Some might argue these outputs are "too rich" and lack a minimalist touch, but it raised a critical question: Is this level of one-shot quality consistent?
The Challenge: Urban Poetry
Instead of optics, I chose a subject of immense architectural and emotional complexity: Invisible Cities by Italo Calvino. The book describes 55 imaginative cities, each serving as a metaphor for a specific human emotion or state of mind.
"When a man rides a long time through wild regions he feels the desire for a city. Finally he comes to Isidora, a city where the buildings have spiral staircases encrusted with spiral seashells..."
Having experimented with city generation back in 2019 using GPT-2, I wanted to see how the current frontier models handled a complex, end-to-end visualization task.
The Prompt
I used a single, straightforward instruction:
Make a three.js (pnpm) visualization of all Invisible Cities by Italo Calvino.
The Contenders: Astra vs. Opus
1. GPT-6 Astra (via Codex)
The result was surprisingly functional. It worked end-to-end, but the "soul" of the design was lacking.
- Performance: 53 minutes of "medium effort."
- Cost: \approx \10$ in API tokens.
- Critique:
- Included "design slop" (excessive comments and concepts that added noise).
- Featured "Captain Obvious" descriptions that felt like accessibility alt-text rather than intentional UI.
- Oddly mimicked the Claude aesthetic (beige backgrounds,
05style numbering).
2. Claude Opus 5.5 (via Claude Code)
I gave Opus 5.5 the same prompt. Despite my instruction to take its time, it finished in 1 hour and 25 minutes. However, this was a result of agentic time dilation.
While the clock said 85 minutes, the 6 parallel subagents combined for roughly .
The Verdict: The result was simply "wow!" Even without using the full six hours I offered, the polish was staggering.
Comparison Summary
| Feature | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|
| Time (Wall Clock) | 53 Minutes | 1 Hour 25 Minutes |
| Compute Effort | Medium | High (6 Parallel Agents) |
| API Cost | \approx \10$ | \approx \74$ |
| Visual Quality | Functional / Noisy | High Polish / "Wow" |
| Style | Derivative | Intentional |
Reflections on the Future of Design
This experiment leaves me with a mix of awe and existential dread. We are entering an era where AI is not just a helper, but a tool that can burn through a significant token budget to produce high-end design.
- Test prompt across multiple models
- Compare cost vs. quality
- Determine the human's role in interactive media
If a model can handle the architectural poetry of Calvino in one shot, what is my place in the creation of interactive media?
Related Insights
🧩 Puzzles and Logic
GPT-6 Astra has shown a massive leap in reasoning. It solved twice as many Baba Is You levels as Claude Fable 5.1. This trend is consistent across:
ARC-AGI-3- Portal (3D gaming)
MazeBench- Obscure puzzles by Przemysław “Psyho” Dębiak.
🍄 Knowledge vs. Wisdom
A recent test involved asking various models "What mushroom is that?" using 360 photos of poisonous species. The contenders included:
- GPT-6 Astra
- Gemini 3.8 Flash
- Claude Fable 5.1
- GPT-5.6
- GLM-5.3-Flash
Pro Tip: Your agent session transcripts are valuable assets. While AI vendors hoard this data, many human teams mistakenly discard their own coding session logs. Keep them!