Real users, real queries, anonymous models.
Two chatbots side by side. You pick the better response. The models’ identities stay hidden until after you vote. Created by LMSYS (UC Berkeley), Chatbot Arena fits a Bradley-Terry model to millions of head-to-head comparisons, producing rankings that track human judgment closely. The prompts are diverse, the voters are real, and the anonymous setup makes it harder to game than fixed benchmarks.