Five models, one prompt, and - for once - a real hands-on playtest on top of the usual line-by-line code read. Part of our ongoing /design benchmark series in Command Code.
The prompt
Create a modern stylish version of the classic offline dinosaur endless runner game (the one that appears in a browser when internet is lost) as a single HTML file using HTML5 Canvas. Use a wide horizontal canvas (800x400)... Graphics should be modern and vibrant... Obstacles include cacti, birds, and modern obstacles like broken robots or barriers... Show high score and current score at the top. Include a nice "No Internet" start screen... Make the movement smooth with good physics, responsive controls, and overall polished addictive feel. Use requestAnimationFrame for 60fps.
(Full prompt in the repo.)
Seen on X
This four-model round was posted with the actual clip and the same hands-on scores used below:
GPT-5.6 Sol vs Fable 5 vs Muse Spark 1.1 vs Grok 4.5 We tested all four using the Chrome Dino game via the /design command....
Scorecard
Design + gameplay, scored 1-10 by hands-on play:
| Model | Feel | Score | Cost | Quality/$ |
|---|---|---|---|---|
| GPT-5.6 Sol | easy to play | 9/10 | $0.53 | ~17.0 |
| Fable 5 | fast | 8/10 | $0.38 | ~21.1 |
| Grok 4.5 | slow but hard | 7/10 | $0.15 | ~46.7 |
| Muse Spark 1.1 | fast and hard | 6/10 | $0.10 | ~60.0 |
| Kimi | not played (code-read only) | - | $0.038 | - |
Which model should you reach for
The lesson from this round is worth keeping regardless of which model you pick: a code read and a hands-on playtest can disagree, and when they do, trust the playtest. If you're building something players will actually touch, budget time to play your own output, not just read the diff. On price alone, Kimi is the value pick here - cheaper than every other build in the set with the strongest technical hygiene on paper. On proven feel, GPT-5.6 Sol is the only one this round where the code read and the played impression fully agreed.
That's easier to check when switching models costs one command instead of a new integration. /design in Command Code runs the same prompt against any model you have access to, so you can playtest a cheaper option before committing budget to a pricier one.
Verdict
GPT-5.6 Sol is the clear winner on actual play - "easy to play," the top score of the four hands-on builds - and this is one of the rare cases where the code read and the hands-on session fully agree: it's the only one of the four with real delta-time physics, a canvas that scales for mobile, and touch controls that can both jump and duck. It's also the most expensive build by far and lands last on quality-per-dollar (~17.0) among the played builds.
Fable 5 plays "fast" and lands second despite being the least technically robust build in the code read - its pace apparently outweighs its gaps once you're actually playing. Grok 4.5 plays "slow but hard," a pacing mismatch where a gentle speed ramp is paired with obstacle placement that still punishes. Muse Spark 1.1 has the best code-side story of the three non-GPT builds but plays worst - technical polish that doesn't translate to feel.
Kimi is the value story on paper: it wasn't hands-on played, but the code read puts it ahead of every other build on technical hygiene - delta-time physics, touch duck, audio, and score persistence - at $0.038, cheaper than all four originals. Whether that translates to "actually fun" the way GPT-5.6 Sol's build did is untested.
Try this yourself
1npm i -g command-code1/model gpt-5.6-sol
2/design Create a modern stylish version of the classic offline dinosaur endless runner game as a single HTML file using HTML5 Canvas with a wide horizontal canvas (800x400)...Full prompt and command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.
