The artwork was generated with GPT Image 2 and GPT Image 1.5 (1.5 isn't as capable, but supports transparent backgrounds). The 3D games also use assets from Poly Haven and KayKit models (https://kaylousberg.itch.io/).
A few observations from the experiment: - A single prompt is often enough to generate a surprisingly playable game. - On average, it took ~10 refinement prompts per game to improve the gameplay and polish. - Canvas/SVG graphics generated directly by the model are generally OK but mediocre; GPT Image (partially prompted with Claude) produced much better visual results.
Related HN discussion: https://news.ycombinator.com/item?id=48513728