Artificial intelligence benchmarking startup Arena has reached a $100 million annualized revenue run rate just eight months after launching its commercial business, highlighting how the race to build capable AI models is creating a lucrative market for independent evaluation and post-training optimization services.
The milestone marks a dramatic evolution for the company, which began as an open-source research project at the University of California, Berkeley, in 2023 before evolving into one of the industry’s most influential AI model benchmarking platforms.
Arena is widely recognized for operating one of the AI industry’s most closely watched crowdsourced leaderboards, where millions of users compare competing AI models by evaluating their responses to identical prompts. The rankings have become an important reference point for developers, enterprises and researchers seeking objective comparisons of leading models from companies including OpenAI, Anthropic, Google, Meta and xAI.
The platform has now accumulated more than 10 million user evaluations, creating one of the world’s largest repositories of human preference data for AI systems.
While the public leaderboard remains free, Arena began monetizing its technology in September with the launch of AI Evaluations. This commercial service provides AI laboratories and enterprise customers with detailed performance analytics based on the vast dataset generated by its user community. Rather than relying solely on standardized benchmark tests, the platform captures how real users judge competing AI models across practical tasks, offering developers richer insight into model quality, reliability and user preferences.
The rapid revenue growth suggests that demand for independent AI evaluation has accelerated alongside the industry’s massive investments in frontier models.
“A lot of people don’t even understand that our business is making any money at all; people still see us as an open source project,” Arena co-founder and Chief Executive Officer Anastasios Angelopoulos told TechCrunch.
The company’s financial performance has accelerated sharply over the past year. When Arena announced a $150 million Series A fundraising round in January at a post-money valuation of $1.7 billion, its annualized revenue stood at approximately $30 million. Less than six months later, that figure has more than tripled to $100 million.
The company refers to the figure as annualized revenue, although Angelopoulos noted that the business model differs from traditional subscription software companies.
Unlike conventional software-as-a-service providers that generate recurring subscription income, Arena charges customers based on consumption. Angelopoulos clarified that while the company uses the term annualized recurring revenue (ARR), “our business is making money from consumption,” meaning customer spending depends on usage levels rather than fixed recurring contracts.
Currently in the intelligence industry, competition has increasingly moved beyond training larger foundation models toward improving their performance through sophisticated post-training techniques.
As frontier AI models become more similar in underlying capability, companies are investing heavily in reinforcement learning, human feedback, evaluation systems and preference optimization to improve response quality and differentiate their products. Independent evaluation platforms have therefore become strategically important because they provide model developers with external measurements of performance across diverse real-world use cases.
Although Arena says it currently has no direct competitor after crowdsourced AI comparison startup Yupp shut down in March, the company competes for spending allocated to AI post-training infrastructure.
Angelopoulos said Arena competes “for the same dollar” as human data-labeling companies, including Mercor, Surge AI and Scale AI, all of which provide training data and human feedback that help AI developers improve model performance after initial training.
The rapid expansion of that broader market indicates growing interest in human-generated evaluation data. Other companies serving the AI training ecosystem have also reported explosive growth.
According to earlier industry reports, Handshake’s annualized revenue from AI training services nearly doubled from approximately $550 million in January to almost $1 billion by April. Mercor similarly surpassed $1 billion in annualized revenue earlier this year after reporting roughly $500 million last September. Those figures suggest that spending on AI infrastructure is increasingly extending beyond chips and data centers to include the human evaluation systems needed to refine increasingly sophisticated models.
Arena has also continued expanding the scope of its benchmarking capabilities. Originally focused primarily on text generation, the platform now evaluates AI systems across multiple categories, including coding, computer vision, image generation, and more complex multi-step agent workflows.
Its recently introduced Agent Mode measures how well AI systems complete long-running tasks requiring planning, reasoning, and sustained execution, capabilities that have become central to the industry’s next generation of autonomous AI agents.
The company traces its origins to academic research at the University of California, Berkeley.
Arena was co-founded by Chief Executive Officer Anastasios Angelopoulos and Chief Technology Officer Wei-Lin Chiang, both former UC Berkeley postdoctoral researchers, alongside renowned Berkeley professor and Databricks co-founder Ion Stoica, who initially advised the research project before it formally incorporated as a company in April 2025. Since then, the startup 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.
As leading model developers spend billions of dollars building increasingly powerful foundation models, the ability to measure, compare, and optimize those systems is emerging as a fast-growing market in its own right. Independent evaluation platforms that combine large-scale human feedback with sophisticated analytics are becoming essential infrastructure for AI companies seeking to improve model performance and win market share.
