Just for fun. I try to keep the number of personal AI accounts low, one or two, but right now I have four, new models are around, and I just wanted to see them coding something non trivial before deciding which one to cancel :D

So I wrote a long(ish) prompt, about 13 lines, with the idea of one-shotting a web game. Even the prompt is not entirely original, as I've seen people on YouTube making very similar games, also to test different AI tools.

I've only published the game made by the winner, you can try to guess who coded this before reading the rest of the post (hint: not me!)

https://guillem.github.io/coral-cove/

(Disclaimer: if you find this long after I published it, the URL above might not work. I won't keep it online forever...)

The system was Bazzite Linux running on one of my test Thinkpads.

Harnesses and models

Antigravity CLI 1.1.7 with Gemini 3.6 Flash High

This was very fast, and also very bad. After about 2 minutes it produced something that did run but most controls didn't work, several elements were missing, and the character went through the rocks, including the sea floor. The enemies were glitching all the time and not pursuing me.

Also the frame rate was dreadfully slow, totally unplayable. In fact this is the first game I tried and I was afraid my lab laptop wasn't powerful enough. But that wasn't the case, the code was just bad.

Grok Build Beta 0.2.112 with Grok 4.5 High

This was the second fastest, it took about 5-6 minutes and produced a game that covered all the basics and definitely worked. 3D models and textures were very basic, but all the game mechanics did work. Enemies did pursue me (in the screenshot you can see the effect when the puffer fish attack me) and shooting bubbles at them worked as defense. I could complete the game by collecting stars and going through the gate.

Kimi Code 0.29.1 with Kimi K3 High

I have mixed feelings about this one :D It took longer, about 25 minutes. It's the only one that chose Python instead of NPM for the local server and it made a much smaller code and deployment, about 2MB total (the others are 50-80MB).

The 3D models are a bit better, more detailed and closer to popular conventions, and the movement is amazing, super smooth. But look at all these "tutorial hints". They didn't go away.

Everything was a bit off. Anyway, I could also complete the game.

Claude Code 2.1.220 with Opus 5 High

This one took about 50 minutes and burned through 33% of my limits for the session and about 10% of the weekly limits. But look at it. It's much closer to an actually playable game. Smoother, more detailed, with better textures... It even made several LEVELS. I played 3, and they're different enough to keep the game interesting for a couple minutes.

Who won?

The game I chose to share with you is the one written by Opus 5.

I created an empty repo on GitHub and I used Claude to commit / push, as well as create GitHub Actions to publish it on the public URL you can test.