4 AI Models Build the Same Walking 3D Character in Godot
2 min read
Build log ·
Can an AI model build a game character that holds together when it walks? Claude Opus 5.5, Sonnet 5.5 and Haiku 4.5 in Claude Code, and GPT-6.1 Sol in Codex, got the same prompt and the same reference picture: a stocky dwarf miner with a braided beard, a helmet lamp, an apron and a pickaxe, built only with Flockbay's modeler, rigged so he can idle, walk and jog. No Tripo, no Meshy, no Blender. Each worked directly in the Flockbay app, on the Godot engine, looked at its own builds, and decided when it was done.
The test
Then all four were filmed in the same studio, under the same lights and lens: a turntable, the front, side and back, the walk and the jog, and close-ups of the hands and the feet.
The prompt
The exact prompt all four got:
Build a game-ready 3D character for a Godot game: a stocky dwarf miner with a thick braided beard, a leather helmet with a lamp on the front, a wool shirt under a leather apron, a wide belt, heavy boots, and a pickaxe in his right hand. Make him look like the reference picture. Build him only with Flockbay's modeler: read flockbay_model_guide, then build with flockbay_build_model. Don't use Tripo, Meshy, Blender or any other 3D tool or service, don't download a model, and don't write mesh code yourself. Give him a skeleton with his bones named by role, so he gets Flockbay's idle, walk and jog. He has to look right from every side, not just the front, and hold together while he walks: his knees, elbows and shoulders bend cleanly, his beard and apron move with him, nothing pokes through, and the pickaxe stays straight and in his hand. Keep him within a triangle budget a game could ship. Look at your own previews, fix what you see, and hand him over when you think he's done.
What was judged
The character against the reference: silhouette, proportions, colours, from every side.
The walk: joints, the beard and apron, the pickaxe, the feet on the floor.
The build: triangles against the budget it set, materials, bones and humanoid roles, height and facing.
How it worked: builds, previews looked at, what it fixed between builds.
What it cost: tokens, calls, time, and the estimated API bill.
In the video
[0:00](https://www.youtube.com/watch?v=jCOTvzumPe4&t=0s) Opus won three, Sol won one
[0:16](https://www.youtube.com/watch?v=jCOTvzumPe4&t=16s) One character this time
[0:41](https://www.youtube.com/watch?v=jCOTvzumPe4&t=41s) The rules: the modeler only
[1:17](https://www.youtube.com/watch?v=jCOTvzumPe4&t=77s) Four models
[1:41](https://www.youtube.com/watch?v=jCOTvzumPe4&t=101s) What a good one needs
[2:10](https://www.youtube.com/watch?v=jCOTvzumPe4&t=130s) What we are testing
[2:57](https://www.youtube.com/watch?v=jCOTvzumPe4&t=177s) Claude Haiku 4.5
[3:41](https://www.youtube.com/watch?v=jCOTvzumPe4&t=221s) Claude Sonnet 5.5
[4:17](https://www.youtube.com/watch?v=jCOTvzumPe4&t=257s) GPT-6.1 Sol
[5:26](https://www.youtube.com/watch?v=jCOTvzumPe4&t=326s) Claude Opus 5.5
[6:32](https://www.youtube.com/watch?v=jCOTvzumPe4&t=392s) All four side by side
[6:49](https://www.youtube.com/watch?v=jCOTvzumPe4&t=409s) None of them put a floor under it
[7:53](https://www.youtube.com/watch?v=jCOTvzumPe4&t=473s) The bill
[8:24](https://www.youtube.com/watch?v=jCOTvzumPe4&t=504s) Who takes it
[9:14](https://www.youtube.com/watch?v=jCOTvzumPe4&t=554s) Tough on purpose, and what's next
The results, measured
| Opus 5.5 | GPT-6.1 Sol | Sonnet 5.5 | Haiku 4.5 | |
|---|---|---|---|---|
| minutes | 10.9 | 22.8 | 4.0 | 4.9 |
| builds | 5 (1 failed) | 10 (1 failed) | 4 | 4 (3 not built) |
| cost on API prices | $2.93 | $0.98 | $0.64 | $0.45 |
| looked at the walk | own renders, no floor | Play and posed, no floor | no | no |
| height | 1.48 m | 1.53 m | 1.45 m | 1.93 m |
| hands, up close | mittens | crude fist (right), mitten (left) | stumps, handle through | sticks |
| hand lower in the walk | ≤ 3.6 cm | ≤ 16.8 cm | ≤ 7.3 cm | ≤ 16.4 cm |
| pickaxe under the floor (side) | never | the whole walk, ≤ 10.6 cm | every other step, ≤ 5.2 cm | 8% |
Costs are estimates at public API prices, read on the day. All four runs went through a subscription.
From the filmed games
