Four models, one prompt, hands-on play instead of a static code read this time. Part of our ongoing /design benchmark series in Command Code.
The prompt
Create a classic Snake game as a single HTML file using Canvas. Grid-based movement, arrow keys or on-screen touch buttons to steer, snake grows as it eats food, game ends on wall or self collision. Neon glow visual style, score display, high score, and a start screen plus game over screen with restart. Make it feel smooth and satisfying to play on both desktop and mobile. Output only the full code.
Models and cost
| Model | $/MTok in/out | LOC | Est. cost |
|---|---|---|---|
| Muse Spark 1.1 (Meta) | $1.25 / $4.25 | 885 | ~$0.06–$0.09 |
| Grok 4.5 (xAI) | $2 / $6 | 597 | ~$0.05–$0.08 |
| GPT-5.6 Sol (OpenAI) | $5 / $30 | 630 | ~$0.40–$0.60 |
| Fable 5 (Claude) | $10 / $50 | 609 | ~$0.44–$0.66 |
Each model was re-rolled 2-3 times before landing on the kept build; cost is a rough floor scaled for the extra attempts, not a metered bill.
Method
This round is judged by hands-on play, not a line-by-line code read: all four builds were played side by side on desktop and mobile, with a quick code check as a sanity pass on a few of the gameplay impressions - not the basis for the verdict.
Scorecard
Design + gameplay, scored 1-10 by hands-on play after re-rolls, on the kept build:
| Model | Score |
|---|---|
| Muse Spark 1.1 | 9/10 |
| Fable 5 | 8/10 |
| Grok 4.5 | 7/10 |
| GPT-5.6 Sol | 7/10 |
Verdict
Muse Spark 1.1 wins. Going in, GPT-5.6 Sol looked like the favorite on recent design form, with Grok 4.5 and Fable 5 also expected to be competitive. Side by side, Muse Spark 1.1 stood out on both looks and moment-to-moment gameplay, at the cost of defaulting hard to a cyberpunk aesthetic rather than a more neutral neon treatment. Fable 5 is a close second - very smooth, held back only by rough turn handling around food pickups. Grok 4.5 is dependable and correct but simply not as polished. GPT-5.6 Sol's sparkle effect needs to be toned down for mobile, where it currently hurts readability.
Cost and quality don't line up neatly here either: Muse Spark 1.1 and Grok 4.5 are both an order of magnitude cheaper than Fable 5 or GPT-5.6 Sol, and Muse Spark 1.1 also won on feel.
Why it's worth running this test yourself
"The favorite going in" lost this round. That's the real argument for testing rather than assuming: a model's reputation on other kinds of work doesn't reliably predict how it'll handle a specific brief like a neon Snake game. The only way to know is to run the same prompt yourself and play what comes back.
/design in Command Code is built for that - one command to switch models, the same session and cost dashboard throughout, so testing a "long shot" model costs you a few minutes instead of a new setup.
Try this yourself
1npm i -g command-code1/model muse-spark-1.1
2/design Create a classic Snake game as a single HTML file using Canvas. Grid-based movement, arrow keys or on-screen touch buttons to steer, snake grows as it eats food, game ends on wall or self collision. Neon glow visual style, score display, high score, and a start screen plus game over screen with restart. Make it feel smooth and satisfying to play on both desktop and mobile. Output only the full code.Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
