Yana Lialuke
GPT-6.1 Sol built a steam train scene in three.js on par with Claude Sonnet 5.5's for $8.33, 5.6x less than Sonnet's $46.28. Sonnet finished 29 minutes faster, in 1h 55m against 2h 24m.
The setup
Both models got one spec and one reference image, and built the scene in chunks and review rounds.
- Models: Claude Sonnet 5.5 and GPT-6.1 Sol
- Sonnet ran as a background agent on a Claude subscription, Sol through OpenRouter
- One scene in three.js, generated entirely in code, with no 3D models, textures or image files
- Each model first wrote a plan and a chunk map, with files under 350 lines and chunks under 150, then built the scene piece by piece
- After every build, a headless browser rendered it next to the reference. The model wrote its own critique and fixed the biggest differences
- Test dates: Nov 6–8, 2026
The time and cost below count only the models' own rounds.
The brief
The subject is a technical one: a steam locomotive on a tall, curved stone viaduct over a mountain gorge. The bar was a scene a viewer could mistake for a game screenshot.
- A green and black locomotive with a tender and carriages
- Rods and pistons that move with the wheels, and smoke puffs in time with the pistons
- A gorge with rock walls, a forest, a river, a waterfall and snowy peaks
- A camera that never stands still
The numbers
Sol's cost is what OpenRouter billed: 57 requests, 2.79M input and 397k output tokens. Sonnet ran on our subscription, so its cost is its tokens at the API list price. Both figures cover only the models' own work.
| Claude Sonnet 5.5 | GPT-6.1 Sol | |
|---|---|---|
| Time | 1h 55m | 2h 24m |
| Cost | $46.28 | $8.33 |
| Plan | 18 files, 29 chunks | 21 files, 31 chunks |
What the models built
In the first comparison, both scenes had the full locomotive with moving rods, smoke and close-ups of the wheels. Sonnet's viaduct was a wall of flat brick. Sol's camera went back to the wheels and the cab after the front view, where it was meant to pull away.
Our take
GPT-6.1 Sol is the value pick: a scene close to Sonnet's for under a fifth of the price, even with fewer rounds and no hints. Sonnet 5.5 was faster, but it cost 5.6x more. Both scenes still fall short of a game screenshot: the rocks, the light and the smoke give them away.
Sources
- Claude Sonnet 5.5 on OpenRouter – the list price we used for the Sonnet estimate
- GPT-6.1 Sol on OpenRouter
- three.js – the 3D library both scenes use
- Spec, rounds, timings and bills: our test, Nov 6–8, 2026
