Description
PacBench is a web gallery that compares how well different AI models and coding harnesses can build a Pac-Man game from one short prompt: create a Pac-Man game in a single HTML page. Each entry is shown as a card that you can open and play, and the gallery can be searched, filtered by model, and sorted by score, cost, time, tokens or release date. Where the harness transcripts allow, cards record wall time, token counts, cost and the size of the generated HTML file. The notes explain how runs were made, for example that some phases used Claude Code pointed at OpenRouter, others used Antigravity with Gemini, and how costs were calculated for particular entries. It is a useful reference for developers curious how coding agents compare on a small, self-contained game task, and the source is on GitHub.…
Highlights
- Playable Pac-Man games generated by different AI models
- Search and filter entries by model
- Sort by score, cost, time, tokens or release date
- Run data with wall time, tokens and HTML size
- Notes on which harness produced each entry
Ratings & Reviews
No ratings yet
Be the first to rate Pac-Bench.
People also visit
Pac-Bench FAQ
What does PacBench measure?
How well a model and harness can recreate Pac-Man from a single prompt asking for a Pac-Man game in one HTML page.
What stats are shown for each entry?
Score, cost, wall time, token counts and HTML file size where the run recorded them; a dash means the number was not recorded.
Which tools were used to generate the games?
The notes say some phases used Claude Code with OpenRouter, others used Antigravity with Gemini, and some entries were written in chat or as API reruns.
Where is the source?
At github.com/jonclegg/pacman-bakeoff.
What are the best Pac-Bench alternatives?
Popular alternatives to Pac-Bench include AI Gallery, TetrisBench and Benchmark Heaven. See all Pac-Bench alternatives

