Seven models, one prompt, a full line-by-line static read of every file. Part of our ongoing /design benchmark series in Command Code.
The prompt
Create a fun endless runner game as a single HTML file using Canvas. Vertical canvas (400x700 or similar portrait mode), character runs automatically forward, swipe or space to jump, down arrow to slide. Collect coins, avoid obstacles like barriers and trains. Nice colorful city visuals with scrolling background, score based on distance, game over screen with restart. Keep it simple but visually appealing. Output only the full code.
Scorecard
1-5 per axis, single sample per model:
| Model | Spec fidelity | Collision | Architecture | Resize robustness | Polish | Offline-safe | Mean | Cost | Quality/$ |
|---|---|---|---|---|---|---|---|---|---|
| Opus 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5.00 | ~$0.37 | ~13.7 |
| GPT-5.6 Sol | 5 | 4 | 5 | 5 | 4.5 | 5 | 4.75 | $0.3265 | ~14.5 |
| Grok 4.5 | 5 | 4 | 4.5 | 3.5 | 5 | 5 | 4.50 | $0.4000 | ~11.3 |
| Fable 5 | 5 | 4 | 4.5 | 3.5 | 4 | 5 | 4.33 | $0.6000 | ~7.2 |
| GPT-5.5 | 5 | 4 | 3.5 | 3 | 4 | 5 | 4.08 | $0.7000 | ~5.8 |
| Gemini 3.6 Flash | 2.5 | 3 | 4 | 2 | 4.5 | 5 | 3.50 | not estimated | - |
A 5.00 mean is not "flawless" - the six axes above predate this run and don't include persistence. Add one and Opus 5 scores a 2 (in-memory best only) against GPT-5.6 Sol's and Gemini's 5; on a 7-axis mean, GPT-5.6 Sol would edge ahead, 4.79 to 4.57.
Which model should you reach for
If depth and game feel are what you're optimizing for, Opus 5 is the clear pick - it's the only model here that renders real 3D instead of layering flat sprites, and the only one with proper frame-rate-independent physics. Just budget five minutes to add localStorage persistence yourself, since it's the one detail every one of these models should have shipped and Opus 5 didn't. If price-to-quality is the priority, GPT-5.6 Sol is the stronger all-around pick specifically because it got persistence right.
Gaps like a missing save-score call are exactly the kind of thing worth catching before you ship, and running the same prompt against a second model is the fastest way to catch them. /design in Command Code makes that comparison a /model swap away.
Verdict
On quality: Opus 5 > GPT-5.6 Sol > Grok 4.5 > Fable 5 > GPT-5.5 > Gemini 3.6 Flash. Opus 5 wins by a category difference, not a points difference - it's the only entry with real projected 3D, frame-rate-independent physics, jump-buffer/coyote-time forgiveness, train-roof standing surfaces, and pause plus blur-auto-pause. The one place it loses is the one that matters for replay value: no persisted high score. Fix that - a handful of lines - and there's no argument left.
On price, Opus 5's ~$0.37 floor sits between GPT-5.6 Sol's $0.3265 and Grok 4.5's $0.40, undercutting Fable 5's $0.60 and GPT-5.5's $0.70, while shipping more than any of them. GPT-5.6 Sol remains the better quality-per-dollar pick (~14.5 vs ~13.7) purely because it also remembered to save the score.
Try this yourself
1npm i -g command-code1/model opus-5
2/design Create a fun endless runner game as a single HTML file using Canvas. Vertical canvas (400x700 or similar portrait mode), character runs automatically forward, swipe or space to jump, down arrow to slide. Collect coins, avoid obstacles like barriers and trains. Nice colorful city visuals with scrolling background, score based on distance, game over screen with restart. Keep it simple but visually appealing. Output only the full code.Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
