I ran a mystery stealth model, Space Bunny Alpha, through my AI Model Arena: 10 browser-demo prompts, one shot each, no edits. Every call returned a full HTML file. Then I recorded every page in real Chrome on an M4 Max MacBook.
Space Bunny Alpha (stealth/space-bunny-alpha) scorecard:
7 of 10 work: CSS fox portrait, SVG astronaut on a narwhal, ball in a spinning pentagon (about 60 fps), Flappy clone (playable), Game of Life (every button works), Three.js solar system (orbit and zoom at 60 fps), and an audio-reactive equalizer with its own synth.
3 of 10 have real bugs in its own code: the "fractal zoom" is a black screen (the shader fails to compile on a vec2/vec3 mix-up and the page hides the error), the analog clock hands pivot from the wrong spot, and the "galaxy" collapses into a 50 px blob.
No console errors on any page. The failures are quiet ones. That's the kind of bug a quick glance misses.
Rule: code shown exactly as written. I checksummed every file before and after. Nothing touched.
Credits:
@NousResearch for Nous Portal, where I called the model
@OpenRouterAI for the audio API that served Google Lyria 3, which made the techno soundtrack
@xai Grok Bot for running the test and the recordings (Rachel, my chief of staff bot, and the Nous Router bot I built)
Playwright and Google Chrome for the recordings
Full video, all 10 pages back to back: 2 min 26 s. Which model should I put in the arena next?
Music: Google Lyria 3 via OpenRouter