WebDev Arena
A leaderboard where humans compare two models' web pages side by side and vote for the better one.
WebDev Arena works by asking people rather than machines. Two models get the same request to build a web page, the results are shown side by side without saying which is which, and a human picks the better one. Thousands of those votes get turned into an Elo rating, the same ranking system used in chess, where a higher number means a model wins more of its matchups.
Its strength is that it measures taste and usefulness, the things automated tests cannot see. Its weakness is the same thing: voters reward what looks good in ten seconds, which is not always what holds up in a real project. Read it alongside the automated benchmarks, not instead of them.