Four models, one prompt, a full static read of 3,399 lines, and a real manual playtest. Part of our ongoing /design benchmark series in Command Code.
The prompt
Create a complete single HTML file for a fun modern car racing game using HTML5 Canvas with a wide canvas (700x700)... realistic car physics including acceleration, deceleration, braking, smooth steering with momentum and realistic drifting when turning at high speed... a smooth, cinematic third-person view directly behind the car.
(Full prompt in the repo.)
Scorecard
Code-read axes (not played):
| Model | Spec fidelity | Physics/feel | Touch/Mobile | Persistence | Polish | Offline-safe | Mean | Cost | Quality/$ |
|---|---|---|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 5 | 4.5 | 4.5 | 5 | 4 | 5 | 4.67 | $0.42 | ~11.1 |
| Grok 4.5 | 5 | 5 | 1 | 5 | 5 | 5 | 4.33 | $0.31 | ~14.0 |
| Muse Spark 1.1 | 3.5 | 3.5 | 4.5 | 5 | 4 | 2 | 3.75 | $0.22 | ~17.0 |
| Fable 5 | 5 | 3 | 1 | 1 | 5 | 5 | 3.33 | $0.80 | ~4.2 |
Manual playtest, design + gameplay:
| Model | Score |
|---|---|
| Fable 5 | 9/10 |
| GPT-5.6 Sol | 9/10 |
| Grok 4.5 | 7/10 |
| Muse Spark 1.1 | 6/10 |
Which model should you reach for
This one's a genuine warning against trusting a code read alone: Fable 5 looks worst on paper - no delta-time physics, no touch support, no persistence - and ties for the best actual play score in the set. If you're shipping something people will touch, playtest it yourself before ruling a model out on a code audit. On price-to-performance, GPT-5.6 Sol is the strongest all-around pick: tied for the top play score at roughly half of Fable 5's cost.
Catching a gap like Fable 5's touch-control miss before a real user does is exactly why it's worth trying more than one model on the same brief. /design in Command Code makes that a /model swap, not a rewrite.
Verdict
On code alone: GPT-5.6 Sol and Grok 4.5 lead (4.67 vs 4.33), Muse Spark 1.1 is third, Fable 5 is last. Played by hand: Fable 5 ties GPT-5.6 Sol for first (9/10), Grok 4.5 takes 7/10, Muse Spark 1.1 takes 6/10. Fable 5's code gaps just don't seem to hurt how it actually plays.
On price, Muse Spark 1.1 ($0.22) and Grok 4.5 ($0.31) stay cheapest even after extra re-rolls; GPT-5.6 Sol ($0.42) and Fable 5 ($0.80) cost more. Weighing play score against price, GPT-5.6 Sol looks like the strongest all-around pick - tied for the best play score at roughly half of Fable 5's cost. Grok 4.5 and Muse Spark 1.1 are the budget picks, at a real drop in how they play.
Try this yourself
1npm i -g command-code1/model gpt-5.6-sol
2/design Create a complete single HTML file for a fun modern car racing game using HTML5 Canvas with a wide canvas (700x700), realistic drifting physics, and a a smooth, cinematic third-person view directly behind the car.Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
