AI Model Leaderboard

Every vote is a pairwise match, and a Bradley–Terry regression over all matches estimates each model's strength on an Elo-like scale: the reference model is pinned at 1500, and a 400-point lead means roughly 10× higher odds of being preferred. The small figure below each score is a 95% confidence range. With Style control on, the regression also holds formatting and length constant, so a model can't climb by looking polished or writing more; off ranks on the raw votes.

The same Bradley–Terry regression, but a "win" is holding your attention longer than another anonymous answer in the same turn — with an answer's position in the list and its length always regressed out. With Style control on, formatting is held constant as well. Ratings sit on an Elo-like scale: the reference model is pinned at 1500, a 400-point lead means roughly 10× higher odds, and the small figure below each score is a 95% confidence range.