Arena AI

Arena AI is a crowdsourced LLM evaluation and leaderboard platform originating from UC Berkeley, formerly Chatbot Arena / LMArena, ranking frontier models by human blind-vote preference. The submitted name "Code Arena" is in fact its code evaluation sub-product (arena.ai/code); the listing name follows the real brand, Arena AI.

Core capabilities

• Human blind-vote battles: the same prompt is answered by two anonymous models, and the user blind-picks the better one before identities are revealed
• Multimodal leaderboards: seven Arenas covering text/code/vision/image/video/search/agent
• Code Arena: two models build runnable web apps in parallel sandboxes; voting requires previewing both first, with three-aspect scoring for functionality/usability/fidelity
• Fullstack evaluation: supports PostgreSQL, authentication, third-party APIs, and one-click deployment to Vercel
• Ranking methodology: Bradley-Terry plus Style Control to remove writing-style bonuses, with the open-source arena-rank package published
• Cost-performance Pareto: shows performance against cost per task

Use cases Model selection and shortlisting by developer/engineering teams, AI researchers, pre-release testing by model vendors, procurement support for enterprises, and ordinary users who want to try frontier models for free; better used as a screening starting point than a production basis.

Unique strengths Complementary to, not a replacement for, Artificial Analysis (automated metrics + price and speed); compared with the discontinued HF Open LLM Leaderboard and SWE-bench, the differentiation is large-scale human blind-vote preference + fully interactive apps + cost-performance Pareto. Note: it shares only the name "Arena" with Design Arena (designarena.ai) and Alpha Arena (nof1.ai); the companies and tracks differ.

Editorial review UC Berkeley lineage, 5M+ monthly active users, 82M+ votes, 400+ models; a general reference cited by media and during model release seasons. The consumer side is entirely free and voting requires no login; /pricing tested as a soft 404. Points deducted for leaderboards being gameable (the Llama 4 private variant incident) and for "preference ≠ correctness"; best used as a shortlist screener. Rating 4/5.