Ai evaluation platform Arena has hit $100 million in annualised run-rate revenue just eight months after launching its commercial service. The company started as a research project at the University of California, Berkeley in 2023.
Its main product is a crowdsourced leaderboard for artificial intelligence models. This site relies on more than 10 million user evaluations. Visitors type a prompt, send it to two different models, and then select which result they prefer.
The leaderboard remains free for public use. Revenue began flowing in September with the launch of Arena Evaluations. This service gives model labs and large enterprises detailed performance analytics drawn from the community.
Growth in income suggests commercial clients are as interested as the evaluators. Many evaluators join the platform to get early access to unreleased models.
“A lot of people don’t even understand that our business is making any money at all; people still see us as like an open-source project,” said Anastasios Angelopoulos, co-founder and chief executive.
Arena reports this figure as ARR, a term usually meaning annualised recurring revenue. Angelopoulos clarified the company charges for consumption. This means the income is not recurring.
There are no direct competitors. Yupp, another crowdsourced model-selection startup, shut down in March. Angelopoulos stated the firm competes for the same dollar as human labelling startups like Mercor, Surge, and Scale AI. These companies help model builders refine their systems after training.
As providers push for better model performance, demand for post-training refinement services keeps rising. In January, Arena announced a $150 million Series A raise at a post-money valuation of $1.7 billion. At that time, annualised revenue was $30 million.
Elsewhere, Handshake’s gross annualised revenue from AI training nearly doubled since January. It climbed from $550 million to almost $1 billion, The Information reported in April. Mercor’s annualised revenue also topped $1 billion earlier this year, up from $500 million in September, according to The Information.
Arena ranks models on tasks including text, coding, vision, and image generation. It also handles complex, long-running workflows through its recently introduced Agent Mode.
Angelopoulos co-founded the company with Wei-Lin Chiang, a fellow postdoctoral student at the University of California, Berkeley. Chiang serves as chief technology officer. Ion Stoica, a renowned professor at the University of California, Berkeley and co-founder of Databricks, also co-founded the startup. He advised the project before it incorporated as a company in April 2025.
Arena has raised a total of $250 million from investors including Felicis, Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, Lightspeed Venture Partners, Laude Ventures, and UC Investments.
What it means
The revenue model shows that companies willing to pay for high-quality human feedback are finding it valuable. The shift from a purely academic tool to a paid data source confirms that model builders need reliable benchmarks to improve their systems.
