LMArena began as the LMSYS Chatbot Arena, a research project from UC Berkeley researchers. You enter a prompt, get answers from two anonymous models, and vote for the one you prefer, and only then are the model names revealed. Votes from a very large number of people are converted into ratings, using a statistical method similar to chess Elo scores, which produces a ranking that reflects real preferences instead of fixed test questions.
The site has grown beyond text chat. There are separate leaderboards for coding, vision, image generation, search, and video, as well as categories such as hard prompts or particular languages. AI labs often test unreleased models there under code names, and a high rank is widely used as a signal in announcements. Users can also chat with specific models directly, and the site publishes some of its data for researchers.
Arena is free to use. The approach has limits: rankings reflect what voters like, which can reward style, length, and formatting over accuracy, and critics have argued that labs can tune models to do well on it. Treat the leaderboard as one signal alongside task-specific tests. Artificial Analysis, Hugging Face’s leaderboards, and Scale AI’s SEAL leaderboards measure models differently. For a quick view of how popular models compare, Arena is the most cited source.