AI Evaluation Startup LMArena Valued at $1.7 Billion in New Funding Round
The Takeaway
- LMArena secured $150 million in new funding, reaching a $1.7 billion valuation.
- Startup provides widely cited AI model rankings for major tech companies.
- The money will help it run AI models and expand its staff.
LMArena, a startup that operates a widely cited ranking of AI models based on their performance, has raised $150 million at a valuation of $1.7 billion, including the new money, according to the company. That’s nearly triple the valuation of its seed funding round, announced in May 2025.
The funding round, co-led by existing investors Felicis and the University of California’s investment arm, will help the company pay for computing power to run AI models it evaluates for customers such as OpenAI, Google, xAI and Microsoft, as well as hire technical staff. Millions of people visit LMArena’s site to rate the models based on head-to-head comparisons, and those ratings inform the rankings.
LMArena was generating revenue at an annualized pace of several million dollars in September. Its recent revenue pace couldn’t be learned. It estimates its “annualized consumption run rate,” a projection of how much revenue it would earn if recent customer usage continued over a year, hit $30 million last month.
The company, started as an open-source project nearly three years ago, says it has more than 5 million monthly users in 150 countries. It’s not clear if those include visitors to its website as well as the people who rate models.
LMArena’s website invites anyone on the internet to ask questions or use models for creations such as images. Those people then pick the best answer between two options, before the site reveals which model generated each output. LMArena tallies the results into a leaderboard for various categories, such as for AI coding, image and video generation.
The startup sometimes hosts models before they have been publicly launched, giving the companies that produce them a way to get early feedback on their models. AI model developers publicly tout their models’ LMArena rankings as the differences between AI models have narrowed across the industry.
“Leading labs are using us because it has been challenging for them to know if their model is good or bad,” said CEO and co-founder Anastasios Angelopoulos.
Some model makers have complained that LMArena’s approach—tapping unpaid Internet users for feedback—is flawed, saying it can be gamed or doesn’t reflect expert opinion. In contrast, LMArena rivals such as data labelling startup Scale AI pays experts such as lawyers or professors to provide feedback on models.
LMArena has argued that random users are often better-positioned to grade the answers to their own questions and that LMArena gets more honest feedback by not paying experts.
“Some of the highest quality evaluation you can get—it’s called ‘golden standard’—is when people vote on the topics they know,” LMArena co-founder Ion Stoica told The Information late last year.
Stoica, the University of California, Berkeley computer science professor who previously co-founded Databricks, launched LMArena as a project called ChatBot Arena with Berkeley graduate students Angelopoulos and Wei-Lin Chiang. The startup was originally funded with grants and donations. It became a for-profit company in May.
The startup is considering expanding its offerings to include using its data to train AI models using reinforcement learning, the training technique that rewards a model for accomplishing certain goals and penalizes it for other behaviors, Stoica previously said. Feedback from people can be used to train models, though with the rise of “reasoning” models, it is increasingly popular to draw some of this feedback from other AI models, rather than from people. Expanding into reinforcement learning related services could help keep LMArena relevant to its customers’ needs.
“Once you become the de facto benchmarking layer, the product naturally expands,” said Peter Deng, a Felicis general partner who co-led the new financing. “The real value comes from deep partnerships with [AI] labs—combining their internal data with comparative external data.”
Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, Lightspeed Venture Partners and Laude Ventures also participated in the most recent funding, which brings the startup’s total funding to more than $250 million.