Four AI coding models, one prompt each, the same reference pictures, high effort. Four tests:
- Build a Lamborghini-style supercar in Blender, from code, from five reference views.
- Build a browser driving game around that car (three.js, one file, touch buttons for phones).
- Rebuild Leonardo da Vinci's aerial screw in 3D and animate it coming apart, part by part.
- Same for a Leonardo gear hoist he drew already exploded: pegged wheels, lantern gears, a winch, a weight.
Gemini 4 Carbon ran in Google Antigravity on my Gemini API key, Claude Fable 5.2 and Claude Opus 5.5 in Claude Code, GPT-6.1 Sol in Codex.
Everything they built is playable here: https://www.kartit.net/games/ai-battle/ (4 games, 4 3D cars, 8 Leonardo animations, keyboard or touch).
The blind vote
Entries were shown as letters in random order. Three AI judges (Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol) scored them blind, and so did I.
| Test | Gemini 4 Carbon | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6.1 Sol |
|---|---|---|---|---|
| Lamborghini, AI jury / me | 8.03 / 6 | 6.43 / 8 | 5.60 / 5 | 6.37 / 4 |
| Aerial screw, AI jury / me | 8.63 / 8 | 7.97 / 9 | 8.07 / 7 | 7.03 / 6 |
| Gear hoist, AI jury / me | 8.20 / 7 | 7.97 / 10 | 7.03 / 8 | 6.80 / 6 |
The AI jury put Carbon first on every judged test. I put Fable first on every test. The driving games were not judged: play them and decide.
Time and cost (all four tests)
| Model | Time | Cost |
|---|---|---|
| GPT-6.1 Sol (Codex) | 39.7 min | $2.11 at API list price |
| Claude Opus 5.5 (Claude Code) | 73.5 min | $16.77 at API list price |
| Claude Fable 5.2 (Claude Code) | 133.1 min | $41.85 at API list price |
| Gemini 4 Carbon (Antigravity) | 122.3 min | $10.17 billed on my API key |
Fable, Opus and Sol ran on my subscriptions, so their dollars are what the same tokens cost through each API (token counts from each tool's own logs, Claude runs checked with ccusage). Carbon's figure is a real bill. List prices on 11 Nov 2026: Carbon $2/$10 per million tokens (doubling on 1 Dec), Fable $8/$40, Opus $4/$20, Sol $2/$10.
What I took from it
- If you judge by "matches the brief", Carbon wins and is cheap until its price doubles.
- If you judge by "which one would I show someone", I picked Fable blind every time, and it was also the slowest and most expensive.
- GPT-6.1 Sol did all four in under 40 minutes for about two dollars.
One run per model per test, so treat it as a snapshot, not a benchmark. Full write-up with every number, the prompts and the FAQ: https://www.kartit.net/blog/gemini-4-carbon-vs-claude-fable-5-2-opus-5-5-gpt-6-1-sol/
Which one would you have picked?
