Pac-Bench – How well can models one-shot a Pac-Man game?
Summary
PacBench is a Show HN-style project that tests how well a model and its harness can recreate Pac-Man from a single prompt, using a browser-based setup and various model backends. It benchmarks one-shot prompts across multiple models, tracks wall time and token usage, and points to a GitHub repository and a downloadable example to reproduce the results.