About ASCII Bench

ASCII Bench measures something benchmarks usually skip: can a language model draw? Every model gets the same prompts and the same instructions, and people vote on the results without knowing which model made which.

How drawings are made

Each model answers every prompt several times through OpenRouter with an identical system prompt asking for plain ASCII only. Replies are cleaned up (code fences stripped, whitespace trimmed) and rejected if empty, too large, or mostly non-ASCII. Multiple samples per prompt mean one lucky or unlucky drawing can't decide a model's fate.

How voting works

You're shown two drawings of the same prompt by two different models, on random sides. You can pick either, call it a tie, or say both are bad. Newer models and closely matched pairs come up more often, so the ranking sharpens quickly. Each matchup can be voted on once.

How the ranking works

Scores come from a Bradley–Terry model fitted to every vote ever cast — the same approach LMArena uses — shown on the familiar Elo scale (1500 = average). Ties and “both bad” count as half a win each way. Because the fit uses all votes at once, the order votes arrived in doesn't matter.

The bars on the leaderboard are 95% confidence intervals from 200 bootstrap resamples. A model's rank is 1 plus the number of models that are better with confidence, so models that can't yet be told apart share a rank.

Privacy

No accounts. A random ID in a cookie keeps your matchups yours, and your IP address is only stored as a keyed hash for rate limiting.

Credits

Built by Jason Botterill. Thanks to Chris (@chatgpt21) and Angel (@Angaisb_) for help with the original version.