Show HN: Pac-Bench – How well can models one-shot a Pac-Man game?

Category: ai-ml

Tags: benchmark, ai-agents, code-generation, pac-man, static-site

Score: 5.5/10 (Innovation: 6, Technical: 5, Documentation: 6, Utility: 5)

Pac-Bench is a benchmark that evaluates how well different AI models and agent harnesses can one-shot a playable Pac-Man game from a single HTML prompt. It collects entries, runs automated playtests and source audits, and presents a scored gallery comparing cost, time, tokens, and output size across models.

Target audience: AI researchers, ML engineers, and developers evaluating LLM coding capabilities

Repository: https://jonclegg.github.io/pacman-bakeoff/ · HTML · 4 stars

View on Hacker News