Arena

Arena (formerly LMArena) operates a community-powered platform for evaluating and comparing frontier AI models. Users submit prompts through a consumer web interface; in Battle Mode the prompt is sent to two anonymous models and the user votes for the better response, with model identities revealed after the vote. Votes feed public leaderboards computed with the Bradley-Terry paired-comparison rating system, which the company also extends with additional features such as style control. Besides Battle Mode, the platform offers Side by Side mode (non-anonymous model selection, votes excluded from public rankings) and Direct Mode (single-model chat without voting); prompts from all modes are collected for research. Arena ranks models across text, coding, vision and image generation tasks, and has introduced an Agent Mode covering complex, long-running workflows.

The platform also serves as a distribution channel for pre-release models: Arena works directly with open-source and commercial model providers to expose unreleased variants under codenames or aliases for community testing, with scores added to the public leaderboard under official names if the model meets the provider's release criteria and Arena's leaderboard policy. The site states that since March 2024 it has helped test proprietary and open-source models from major labs and smaller teams.

Arena publishes open datasets and research. It describes itself as having open-sourced the largest repository of organic human preferences on generative models, distributed via HuggingFace and explorable through Arena Explorer, and lists research papers including the original Chatbot Arena paper (ICML 2024), LMSYS-Chat-1M, LLM-as-a-judge (NeurIPS 2023), Prompt-to-Leaderboard, Arena-Hard, RouteLLM, the Vision Arena dataset, Search Arena and work on adversarial attacks on leaderboards and over-refusal benchmarking.

Founding story

The project was created by researchers at UC Berkeley. Anastasios Angelopoulos, now co-founder and CEO, and fellow UC Berkeley postdoctoral researcher Wei-Lin Chiang, now CTO, co-founded the company; UC Berkeley professor and Databricks co-founder Ion Stoica advised the project before it was incorporated as a company in April 2025 and is also a co-founder.

Business model

A free consumer-facing evaluation and chat product generates human preference data at scale; that data is monetized through a paid AI Evaluations service sold to enterprises, model labs and developers. The company states this evaluations business is how it sustains itself financially.

Revenue comes from the AI Evaluations service, introduced in September, which provides deep-dive performance analytics drawn from Arena's evaluator community. The CEO states customers are charged on a consumption basis rather than on recurring subscriptions, so the reported ARR figure reflects annualized, non-recurring consumption revenue.

Traction

TechCrunch reports Arena reached $100 million in annualized run-rate revenue about eight months after launching its commercial service, up from $30 million annualized in January 2026. The leaderboard is generated from over 10 million user evaluations, and the company's own site claims tens of millions of users. Total funding reported is $250 million.

Latest developments

As of June 2026, Arena reported $100 million in annualized run-rate revenue eight months after commercial launch, and had recently introduced Agent Mode for evaluating complex, long-running workflows. Rival crowdsourced model-picking startup Yupp shut down in March 2026.