Seven models, one prompt, a full static read of all 4,754 lines plus an automated headless smoke test. Part of our ongoing /design benchmark series in Command Code.
The prompt
Create a retro pixel art horizontal space shooter game as a single HTML file using Canvas. Vertical or wide canvas, small orange spaceship that can move up, down, left, right. Smooth movement with slight inertia. Shoot with spacebar, collect glowing blue gems for points, avoid obstacles. Starry black background with scrolling cave walls (top and bottom). Simple enemies or obstacles, score display, lives, game over screen. Classic 8-bit/16-bit pixel style with nice glow effects on gems and ship thrusters. Output only the full code.
Scorecard
| Model | Spec fidelity | Collision | Architecture | Resize | Polish | Offline-safe | Mean | Cost | Quality/$ |
|---|---|---|---|---|---|---|---|---|---|
| Grok 4.5 | 5 | 5 | 5 | 4 | 5 | 5 | 4.83 | $0.20 | ~24.2 |
| GPT-5.6 Sol | 5 | 4 | 5 | 5 | 5 | 5 | 4.83 | $0.45 | ~10.7 |
| Fable 5 | 5 | 3.5 | 5 | 5 | 4.5 | 5 | 4.67 | $0.65 | ~7.2 |
| GPT-5.5 | 5 | 4 | 4 | 5 | 3 | 5 | 4.33 | $0.35 | ~12.4 |
| GLM 5.2 | 5 | 4.5 | 5 | 4.5 | 5 | 2 | 4.33 | $0.03 | ~144 |
| DeepSeek V4 Pro | 3.5 | 2 | 3.5 | 1 | 3.5 | 2 | 2.58 | $0.00017 | ~15,176 |
Which model should you reach for
If price is the deciding factor and you can live with patching in collision code yourself, DeepSeek V4 Pro's quality-per-dollar is hard to argue with. If you want a complete, shippable game with no follow-up prompt, Grok 4.5 and GPT-5.6 Sol are both a safer default - they tie for the best score in the set and neither has a defect that needs fixing before you'd hand it to a player.
This is exactly the kind of tradeoff that's hard to see from a single vendor's own benchmark page. Running the identical prompt across seven models in one CLI - same session, same cost tracking, same taste profile carried between them - is what makes a comparison like this possible in the first place. That's the whole premise of /design in Command Code: pick the model that's actually right for the brief, not the one you happened to start with.
Verdict
On quality alone, Grok 4.5 and GPT-5.6 Sol tie for best in the set. Grok wins by shipping the only full progression loop - health, scaling bullets, leveling, and a boss fight. GPT-5.6 Sol has no progression layer but is the only build with true frame-rate-independent physics and no network-font leak.
Price reshuffles the ranking hard. GLM 5.2's $0.03 buys 93% of Fable 5's quality score at 4.6% of the cost. DeepSeek is another ~176x cheaper than GLM and lands at 55% of its quality - but this round's defect (non-lethal cave walls) is a core-mechanic gap that needs real collision code, not a five-minute CSS fix like the resize bug this repo caught from DeepSeek in the flappy-bird benchmark. Grok 4.5 sits in the sweet spot: cheaper than Fable 5, GPT-5.5, or GPT-5.6 Sol while tying the best quality in the set, with no fatal gaps - the strongest overall pick this round on both axes.
Try this yourself
1npm i -g command-code1/model grok-4.5
2/design Create a retro pixel art horizontal space shooter game as a single HTML file using Canvas, with scrolling cave walls, glowing gem pickups, and a classic 8-bit/16-bit pixel style.Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
