Arena

Arena (formerly LMArena) is an open AI model benchmarking platform. It provides an ‘arena’ where users can anonymously compare the responses of different AI models (such as GPT, Claude), and generate an aggregated leaderboard reflecting real-world performance through voting.

In Battle Mode, users submit a query or prompt and the system randomly selects two anonymous AI models to generate responses in parallel. Users vote for the better answer based on quality, and votes affect the models’ Elo scores and leaderboard rankings.

According to public information, the core evaluation and comparison features on Arena are currently freely accessible to users. You can experience and test the integrated AI models on the platform.

The platform uses anonymous battles so voters don’t know model identities to reduce brand bias. An Elo scoring system processes the large volume of votes, and all evaluation data and rankings are publicly auditable.

Arena offers multi-domain evaluations, including text dialogue, visual understanding, image generation, video generation, code programming, web development, and search enhancement, covering the capabilities of mainstream models.

According to the platform’s policy, user input may be processed by third-party AI models and could be disclosed to the respective AI providers and publicly shared to support community development and AI research. Users are advised not to submit sensitive or personal data.

Leaderboards are dynamically updated based on ongoing community votes. Each specialized leaderboard (e.g., Text, Vision) typically shows the most recent update time, such as 'updated 1 day ago', indicating timely rankings.

Traditional benchmarks use fixed standardized test items. Arena emphasizes evaluation based on real user tasks and subjective judgments, reflecting model performance in real-world scenarios through a large volume of anonymous votes and comparisons.