Disposable model test · 25 September 2026
Three models.
One UFO.
I gave Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna the same loose brief: make a small one-button UFO parking game in Three.js.
Three guesses at the same tiny game
These are the actual deployed opening screens, not mock-ups.
The large frame is Opus because it is the one that made me want another round. Open any frame to play the full page.
“The Opus one is ridiculously fun.”
That is the useful result here. Opus added patrols and hardened the levels as the round went on. It stopped feeling like a demo of movement and started feeling like a little game with somewhere to go.
Sol is more deliberate: a flight-ops screen, a landing computer and a clearer systems readout. Luna is the quickest read: quiet, clean and cheap.
What the run actually did
A small brief, with room to tinker.
Each run got the same bundled Three.js file, the same thirteen tools, xhigh reasoning, a 128,000-token ceiling and a writable project folder. The shell ran in a network-disabled Node Docker container.
Media was optional. Opus requested one music loop and two voice clips. Sol and Luna chose code-generated visuals and sound. None requested image, SFX, video or Blender output.
One run each. The checks cover syntax, local references, Three.js loading and input handlers. They do not prove browser feel, accessibility or a fair model ranking.