Plattform, auf der Nutzer über anonyme Modellantworten abstimmen, um Rankings zu erstellen
KI-Benchmarks KI-Modell-Benchmarks arena.ai
What is LMArena
LMArena (now called Arena, and originally Chatbot Arena) is a model evaluation platform created by researchers at UC Berkeley. You enter a question, the platform shows answers from two anonymous models, you vote for the better one, and models are ranked from large numbers of votes. Chatbot Arena launched on 24 April 2023, moved to its own domain, lmarena.ai, in September 2024, and in 2025 became a company, Arena Intelligence Inc.
What it can do
- Anonymous head-to-head voting: compare two models' answers, and learn which models they were only after voting
- Leaderboards ranking models by human votes, across text, images and other arenas (image support was added in June 2024)
- Try many models for free: chat directly with the latest models from different companies
- Pre-release testing: companies often test unreleased models here under code names — DeepSeek tested prototypes months before R1 drew attention, and OpenAI's GPT-5 appeared as "summit"
Pricing
Free for users.
Who it's for
- Anyone who wants to know which model is most useful right now
- Developers and companies who want real-world preference data when choosing a model
- People who want to try new models for free
Good to know
- Research has identified limitations in its methodology. In April 2025, the version of Meta's Llama 4 Maverick on the leaderboard differed from the publicly released version, causing controversy, and LMArena changed its policies.
- Votes reflect user preference and are influenced by answer style and length, so they aren't the same as accuracy; combine rankings with tests on your own tasks.
- Companies including OpenAI, Google DeepMind and Anthropic supply models to the platform, which raised a US$200 million Series B at a US$3.1 billion valuation in October 2026.
Similar tools
Compare the Open LLM Leaderboard, OpenCompass and SuperCLUE; for academic benchmarks, see MMLU and C-Eval.