Pac-Bench

Pac-Bench evaluates how well AI models can generate a complete Pac-Man game from a single prompt requesting a playable game in HTML. Each model receives one opportunity with no follow-up requests or corrections allowed.

The benchmark tests one-shot code generation capabilities across different AI models using an identical prompt: 'Create a Pac-Man game in a single HTML page.' Results showcase real-world performance differences in handling complex, multi-feature requests without iterative refinement.