Yana Lialuke
Claude Opus 5.5 built the best casino heist cutscene of four models, for about $41, less than half of Claude Fable 5.1's ~$87. GPT-6.1 Sol did it for $10.17 and GPT-6.1 Terra for $2.17. Every character, set and camera move was built from code, and none of the four got close to the reference frames we gave them.
The setup
All four models got one brief: a 35-second heist cutscene in 7 shots, built entirely in code.
- Models: Claude Opus 5.5, Claude Fable 5.1, GPT-6.1 Sol, GPT-6.1 Terra
- Opus and Fable ran as background agents on a Claude subscription, Sol and Terra through OpenRouter
- Characters modeled, rigged and animated by each model with Python scripts in Blender
- The scene, lights and cameras in three.js, no images or stock 3D models
- Test dates: Oct 28–29, 2026
Each model first wrote a plan and split the code into chunks. Then it built the characters, the sets and the scene, and fixed them in review rounds on its own rendered frames.
The brief
The crew is a leader in a suit and four masked robbers. The masks are white porcelain, black-and-gold with a glowing LED grin, chrome, and red with horns. The video should look like a cutscene from a 3D game, with fast cuts.
- Smile: a close-up of the leader's grin with a gold tooth, then a fast pull-back into the van
- Arrival: the van brakes at the casino and the crew jumps out
- Entry: the glass doors shatter, shot from three angles
- Slots: the crew runs at the camera between slot machines
- Dance: the leader dances among stacks of cash
- Bag: gloved hands stuff money into a bag
- Hero: a slow-motion leap over a roulette table
How the brief grew
Our first test on Opus built everything inside three.js, and the people came out like mannequins. So we moved the characters to Blender, as in our Opus 5.5 Halloween test. After the base build we ran more rounds on top:
- Fix rounds on the biggest problems in each model's own frames: faces, the van, the bag shot
- Two rounds against 8 reference frames we generated with Gemini 3.1 Flash Image
- A sixth crew member: a clown in a purple coat and a rubber mask, with a new ending
Opus, Fable and Sol got three fix rounds after the base build. Terra joined later and went straight to the reference rounds. The last three rounds were the same for everyone.
The clown ending
In the new ending, the leader's leap cuts at 32.4 seconds to the clown. He pulls off his rubber mask and laughs at the camera. The last shot is his card dropping onto the roulette felt.
The numbers
Sol's and Terra's costs are what OpenRouter billed. Opus and Fable ran on a subscription, so their costs are estimates from their real token usage at the API price. The totals cover every round, from the base build to the clown.
| Model | Time, all rounds | Cost, all rounds | Clown round alone |
|---|---|---|---|
| Claude Opus 5.5 | 3h 18m | $40.55 | $5.09 |
| Claude Fable 5.1 | 3h 12m | $86.93 | $11.27 |
| GPT-6.1 Sol | 2h 45m | $10.17 | $1.86 |
| GPT-6.1 Terra | 58m 50s | $2.17 | $0.24 |
Terra wrote its whole base scene, 51 chunks of code, in 34 minutes for $0.89, and it ran on the first try. Our full review of Terra is in GPT-6.1 Terra: half the price of Sol. Fable cost 2.1x more than Opus and 40x more than Terra. The reference images cost $1.48 in all and are not counted for any model.
Four heists
- Claude Opus 5.5 – the most complete, game-like picture: a glossy van, a gold entrance, the crew running at the camera. Its clown reads best, with the mask in his hand. The sleeves are plain tubes and the neck is too long
- Claude Fable 5.1 – the weakest start: an egg-shaped head with a slit for a mouth. It improved the most over the rounds, but its clown's coat breaks into spikes at the hips
- GPT-6.1 Sol – the best grin and gold tooth of the four from the first build, but the bodies look like inflated jumpsuits. Its clown has the right face on a tactical body
- GPT-6.1 Terra – the fastest and cheapest, close to Sol in quality. Its clown reads well at the slots, but the reveal is broken: he lies behind a counter and the mask hangs in the air
Faces and bodies are where all four hit their limit. Next to the reference frames, every crew still looks like a set of mannequins.
Our take
Opus 5.5 gave the best result for less than half of Fable 5.1's price, and Fable looks nerfed on this test. GPT-6.1 Terra is the budget pick: a full heist with a clown for $2.17 in under an hour. None of the four can yet build characters from code that match a reference frame.
Sources
- Claude Opus 5.5 on OpenRouter – the list price we used for the Opus estimate
- GPT-6.1 Sol on OpenRouter
- GPT-6.1 Terra on OpenRouter
- three.js – the 3D library all four scenes use
- Blender Python API – what the characters were modeled with
- Brief, rounds, timings and bills: our test, Oct 28–29, 2026
