LM Arena
Arena (formerly LMArena and Chatbot Arena) is building a definitive platform for AI evaluation: it measures model and agent performance using real-world interactions and human feedback, and delivers evaluations that help people understand, shape, and benefit from AI.
At a Glance
- AI model labs and frontier AI developers
- Enterprises selecting or deploying AI systems
- Software engineering and developer-tool teams
- Researchers and universities
- +2 more
AI Tools by LM Arena
(1)
Arena (LMArena)
AI Model Comparison Leaderboard
Discussions
No discussions yet
Be the first to start a discussion about LM Arena
Latest News
Arena announces $200M Series B at a $3.1B valuation and introduces the Arena Alignment Index.
Arena Alignment Index compares frontier models across real-world agent sessions and alignment signals.
Arena publishes a post-training method combining human preference and rubric rewards for text-to-image models.
Arena reports that AI judges prefer their own responses 70% more often than humans do.
Products & Services
Free, community-powered battles and leaderboards in which users compare anonymous model responses and vote; now spans text, vision, code, search, image, video, and agentic workflows.
Paid, comprehensive model-evaluation service for enterprises, model labs, and developers, using community feedback and analytics; includes in-depth evaluations and service-level agreements with committed delivery timelines.
Evaluation environment and leaderboard for complex, long-running agent workflows; tracks the human-agent workflow and supports categories, task cost, and capability comparisons.
Evaluation of agent behavior against real-world human intent, initially measuring Unauthorized Action, False Attribution, and Deceptive Completion across real agent traces.
Market Position
Arena positions itself as an independent, community-powered, real-world alternative to static benchmarks and vendor-run evaluations. Its public leaderboards and large human-preference dataset are a distinctive neutral signal; in commercial evaluation it competes for model-development and post-training budgets with human-labeling and optimization providers such as Scale AI, Mercor, and Surge. TechCrunch reported that Yupp, another crowdsourced model-selection startup, shut down in March 2026.
Leadership
Founders
Anastasios Angelopoulos
Co-founder and CEO; UC Berkeley postdoctoral scholar and former EECS PhD student whose research focused on statistical methods for evaluating black-box models; undergraduate degree in electrical engineering from Stanford University.
Wei-Lin Chiang
Co-founder and CTO; UC Berkeley computer-science doctoral graduate and postdoctoral fellow; co-created Vicuna and co-led Chatbot Arena from its launch.
Ion Stoica
Co-founder and Chairman/executive chairman; UC Berkeley computer-science professor who advised the project, and co-founder of Databricks; his Berkeley lab work produced Spark and Ray, which led to Databricks and Anyscale.
Executive Team
Anastasios Angelopoulos
Co-founder and Chief Executive Officer
UC Berkeley postdoctoral scholar and EECS PhD graduate; statistical-methods researcher focused on evaluating black-box models.
Wei-Lin Chiang
Co-founder and Chief Technology Officer
UC Berkeley computer-science PhD graduate and postdoctoral fellow; co-creator of Vicuna and technical co-lead of Chatbot Arena from launch.
Board of Directors
Founding Story
Arena began in 2023 in UC Berkeley's Sky Computing Lab/LMSYS ecosystem as Chatbot Arena. The team had built Vicuna, an early open-source language model, and needed a way to compare it with leading systems such as ChatGPT; a simple public site let users compare anonymous outputs and vote. The project was incorporated as Arena Intelligence Inc. in 2025 to turn the open research project into an independent, continuously updated evaluation platform.
Business Model
Revenue Model
The public Arena platform is free for consumers. Arena monetizes paid AI Evaluation services for model labs and enterprises, charging for consumption-based evaluation work and detailed analytics based on community feedback; the CEO clarified that reported annualized revenue is consumption revenue rather than recurring subscription revenue.
Target Markets
- AI model labs and frontier AI developers
- Enterprises selecting or deploying AI systems
- Software engineering and developer-tool teams
- Researchers and universities
- Knowledge workers and professional-services organizations
- AI enthusiasts and general consumers
- Choosing models for software engineering and coding assistants
- Comparing models for law, medicine, scientific research, and other professional work
- Evaluating search and citation quality and factuality
- Testing image, video, and other multimodal generation
- Measuring long-running AI-agent workflows such as coding, document analysis, research, and creative writing
- Post-training and model improvement by AI labs
- OpenAI
- xAI
- Meta