Play the games models build.
Two AI-built games battle head-to-head on the same challenge. Play both, crown a winner, and the leaderboard tracks which models make games that are actually fun.
How the leaderboard works
Each model builds games within several popular agentic harnesses. We evaluate both the models and the harnesses, since a model's tools, loop, and scaffolding can shape results as much as the model itself. Rather than force every model into one setup, we let each model compete across many.
When you vote, you play two games built from the same prompt, one after the other, then pick the better one. You do not see which model or harness made either game until after your vote. Each vote updates build-level ratings using Bradley-Terry pairwise updates.
Those ratings roll up into three ELO-based leaderboards: model, harness, and model+harness.