How it works
A play-and-vote arena for AI rebuilding the history of games
Every frontier model can build Pong, so testing them is boring. The interesting question is the efficiency frontier: when a new small, cheap, fast model ships, how well can it recreate a real game from a single prompt? arcade-bench answers that with people, not a script.
1
One locked prompt, never tuned
Each model gets the same single sentence per game — only the game name changes. This is a benchmark of models, not of prompting.
Create {game} as a single self-contained HTML file that runs in a browser.2
A single self-contained HTML file
The required output is one file — inline JS and CSS, no build step, no external assets, no network. That single-file rule is the permanent spine: it keeps every entry safe, instantly playable, and fair. One shot, fixed token budget, no human iteration. A build that returns broken code competes anyway and earns the loss — that's signal, not noise.
3
Served in a strict sandbox
Because the site runs code written by AI models, every artifact renders inside a strict sandboxed
<iframe>: scripts only, no same-origin access, no network. That is the one hard security requirement, and it is non-negotiable — model-written JavaScript can never touch the site, your account, or the network.4
Pairwise voting
You see two recreations of the same game side by side, with the original described for reference, identities hidden. You play both and pick the better one — or call a tie or both bad. Pairwise comparison is far more reliable than asking for a 1–5 score, and it is what makes the rating math work. Anyone can play; casting a vote takes a quick sign-in so the human grading stays honest.BUILD AvsBUILD Bmodel hidden
5
Elo → Bradley-Terry ratings
Votes feed a rating — Elo to start, Bradley-Terry as volume grows — the same family of methods LMArena uses for LLMs. Every rating ships with a confidence interval so a five-vote model never reads as equal to a five-hundred-vote one. The leaderboard ranks models overall, per game, and per cost tier, with the headline view being rating per cent spent.