Eight AI models built 24 playable 3D games in a cost test
tests whether AI models can keep space, time, and cause and effect consistent inside a 3D game. Instead of using fixed questions, eight models received the same three assignments: an arena combat game, a physics puzzle, and a racing game. They produced 24 browser-playable 3D games in total.
The tested group included GLM 5.2, Grok 4.5, and models from OpenAI and Anthropic. More were planned, but the cost of running them made broader testing difficult. The full run cost about $1,390, including roughly $756 for three Fable runs.
One Fable physics-puzzle run alone cost almost $491 and took nearly nine hours, while all three runs together cost about $108.
Key points
- Eight models completed the same three game assignments, producing 24 playable 3D games.
- The test checks spatial layout, over time, and whether actions have sensible consequences.
- The complete experiment cost about $1,390.
- Three Fable runs cost about $756; one physics-puzzle run cost nearly $491 and took almost nine hours.
- All three runs cost about $108 combined.