Before it
Models were ranked on static benchmarks that leaked into training data.
The problem
Benchmark scores stopped predicting which model people actually preferred.
What it introduced
Anonymous head-to-head comparisons from the public, aggregated into an Elo-style rating.
Why it matters
It became the field's de facto leaderboard, and exposed how weakly benchmark scores track usefulness.
What modern models inherit
Human preference as the headline ranking, with all its noise.
What came next
Style controls and category breakdowns as the arena's own biases became apparent.