Ten models, one prompt, one shot each, no follow-up corrections. This round of our ongoing /design benchmark series in Command Code is a neon cyberpunk brick breaker.
The prompt
Create a complete single HTML file for a neon brick breaker game using Canvas. Neon cyberpunk aesthetic with glowing bricks, mouse or touch to move paddle, ball with trail effect, particle explosions on brick break, score, lives, start and game over screens. Make it visually striking with glow effects and smooth gameplay. Output only the full code.
Models and cost
| Model | Est. cost | LOC |
|---|---|---|
| Opus 5 (Anthropic) | ~$0.26 (floor) | 1,243 |
| GPT-5.6 Sol | ~$0.52 | 768 |
| Fable 5 (Claude) | ~$0.60 | 546 |
| GPT-5.5 | ~$0.35 | 400 |
| Gemini 3.6 Flash | ~$0.45 (~5 prompts) | 1,311 |
| Kimi K3 | ~$0.10 | 953 |
| GLM 5.2 | ~$0.03 | 592 |
| DeepSeek V4 Pro | ~$0.0017 | 575 |
| Qwen 3.8-27B | ~$0.005 | 1,311 |
| Laguna S 2.1 / Ling 3.0 Flash | not estimated | 931 / 669 |
Cost figures are one-shot floors from output bytes, not metered invoices - treat them as a lower bound, not a bill.
Method
Only opus-5/index.html got a full line-by-line read this round, cited by line number throughout. The other nine builds were checked mechanically - grepped for external URLs, localStorage, devicePixelRatio, delta-time, mute/pause, substepping, and glow technique (shadowBlur vs. real compositing). That's factual but shallower than a full read, which is why only Opus 5 gets a per-axis craft score below - assigning 1-5 scores to code that wasn't fully read would be inventing precision.
Verified feature matrix
Presence/absence, mechanically checked across every build:
| Build | Offline-safe | Best score saved | DPR-aware | Delta-time | Substepped collision | Mute | Pause |
|---|---|---|---|---|---|---|---|
| Opus 5 | ✗ (fonts) | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ |
| DeepSeek V4 Pro | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Fable 5 | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ |
| GLM 5.2 | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ |
| GPT-5.5 | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| GPT-5.6 Sol | ✓ | ✗ | ✓ | ✓ | ✓ | ✗ | ✗ |
| Kimi K3 | ✗ (fonts) | ✓ | ✓ | ✓ | ✗ | ✗ | ✓ |
| Gemini 3.6 Flash | ✗ (fonts) | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ |
| Laguna S 2.1 | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ |
| Ling 3.0 Flash | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
Why this comparison is worth running yourself
The result that stands out here isn't "which model wins" - it's that the strongest build in the entire set still shipped with a real gap (no saved high score), while three much cheaper models got that detail right. No single model in this table is safe to assume "complete" without checking. That's the case for running your own prompt across a few models before you commit to one, rather than trusting a single vendor's demo reel.
/design in Command Code is built for exactly that: the same prompt, any model, one CLI, with cost tracked per run so you can see the price-to-quality tradeoff for yourself instead of taking someone else's word for it.
Scorecard: Opus 5
| Axis | Score | Basis |
|---|---|---|
| Spec fidelity | 5 | Every prompt clause present |
| Glow / visual craft | 5 | Only real two-pass bloom in the set, with a fallback |
| Collision | 5 | Substepped, correct normals, degenerate-case handled |
| Architecture | 5 | Sectioned, delta-time loop, responsive re-layout |
| Polish / features | 5 | 8 levels, 4 power-ups, combo system, 10 SFX, mute, pause |
| Offline-safe | 2 | Three Google Fonts <link> tags |
| Mean | 4.50 |
Verdict
Opus 5 is the strongest build in this directory on every axis that involves rendering or physics, and it fails the one constraint this benchmark applies to every demo. Both facts are load-bearing - if you want the best-looking brick breaker in the set, it's this one, after you inline the fonts and add four lines of localStorage.
Gemini 3.6 Flash took the more interesting side path: its one-shot output wasn't playable and needed 4-5 follow-up prompts, but what it landed on has a genuinely unique drop system - falling coin power-ups, one that widens the paddle, one that boosts ball speed. Opus 5's drop system is the only comparable one, and it arrived in a single shot.
On price, Opus 5's $0.26 floor undercuts GPT-5.6 Sol ($0.52), Fable 5 ($0.60), and GPT-5.5 ($0.35) while shipping more than any of them - though DeepSeek V4 Pro ($0.0017) and GLM 5.2 (~$0.03) remain in a different cost class entirely for anyone optimizing purely on spend.
Try this yourself
1npm i -g command-code1/model opus-5
2/design Create a complete single HTML file for a neon brick breaker game using Canvas. Neon cyberpunk aesthetic with glowing bricks, mouse or touch to move paddle, ball with trail effect, particle explosions on brick break, score, lives, start and game over screens. Make it visually striking with glow effects and smooth gameplay. Output only the full code.Swap /model to compare against any other model you have access to. Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
