Arena AI

Arena AI uses a Bradley-Terry scoring system derived from anonymous pairwise battles. Users submit prompts to two hidden models, vote on the better response, and these votes are aggregated into Elo-like scores with confidence intervals, weighted to account for vote frequency imbalances across models.

Arena Max is an intelligent model router that analyzes each user prompt and routes it to the most capable model for that specific task. Powered by 5M+ community votes, Max achieved the #1 overall Arena score by combining the strengths of multiple frontier models with latency-aware optimization.

Arena AI operates 10+ specialized arenas including Text, Code, Vision, Document, Text-to-Image, Image Edit, Text-to-Video, Image-to-Video, Video Edit, Search, and Expert. Each arena has sub-categories like WebDev (HTML and React), Occupational Fields, and difficulty tiers for granular model evaluation.

Yes. Major AI labs including OpenAI, Google DeepMind, Anthropic, and DeepSeek have submitted pre-release models anonymously to Arena for community evaluation before public launch. Notable examples include DeepSeek R1, OpenAI GPT-5 (codename summit), and Google Gemini's Nano Banana image model.

Yes. User prompts, votes, and conversation data are shared with third-party AI providers and may be published publicly as part of Arena's open research mission. Arena has open-sourced the largest repository of organic human preferences on generative models. Users should avoid submitting sensitive personal information.