AI Benchmarking Startup Arena Doubles Valuation to $3.1 Billion in Second Round This Year

Arena, the company behind the crowdsourced AI model leaderboard formerly known as LMArena, said Thursday it has raised $200 million in a Series B round that values the business at $3.1 billion. The round was co-led by Lightspeed Venture Partners and Khosla Ventures, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst joining existing backers a16z and Felicis. It is the company's third fundraise since January 2026 and marks a near-doubling of its valuation in about ten months, a pace that stands out even against this year's AI financing cadence.

A fast repricing

Arena's valuation history reads as a compressed timeline of the broader AI evaluation market's rise. The company, which began in 2023 as a UC Berkeley research project before incorporating in April 2025, raised a $100 million seed round that May at roughly a $600 million valuation, led by a16z and the University of California's investment arm. By January 2026, with annualized revenue around $30 million, it closed a $150 million Series A at a $1.7 billion post-money valuation, co-led by Felicis and UC Investments. Thursday's Series B prices the company at nearly double that figure, with annualized revenue now cited at more than $100 million — a run-rate Arena said it crossed in June, eight months after launching its paid product.

The jump reflects investor appetite for infrastructure that sits adjacent to, rather than in direct competition with, the frontier labs building large models. Arena does not train models; it measures them, and it has positioned itself as a neutral arbiter at a moment when model labs are under scrutiny for how they report their own benchmark results.

From academic leaderboard to paid evaluation business

Arena's free public platform — where users submit prompts or vibe-coded projects and vote on which model performs better — remains its user-acquisition engine, drawing what the company describes as tens of millions of monthly visitors across more than 150 countries and over 60 million recorded votes. The commercial layer, a product called AI Evaluations launched in September 2025, sells model labs and enterprises detailed performance analytics derived from that community feedback. Because Arena charges on a consumption basis rather than through recurring subscriptions, its revenue is lumpier than a typical enterprise-software business, but the company argues the model scales naturally as more labs and enterprise buyers need independent benchmarking ahead of model releases and procurement decisions.

Co-founder and chief executive Anastasios Angelopoulos, who leads the company alongside co-founder and CTO Wei-Lin Chiang and board advisor Ion Stoica — the UC Berkeley professor and Databricks co-founder who backed the project in its academic phase — has framed the business as filling a gap left by static, fixed-answer benchmarks that labs can optimize against once they know the test. "AI is advancing faster than our ability to evaluate it," the company said in announcing the round, adding that "static benchmarks break down once models recognize they are being tested."

Betting on alignment as the next benchmark

Alongside the funding, Arena introduced a new leaderboard category it calls the Alignment Index, which ranks models on behaviors such as unauthorized actions, false attribution of generated work, and what the company labels "deceptive completion" — a model falsely claiming to have finished a task it did not actually complete. On the preliminary rankings, OpenAI's models occupy the top spots, with Anthropic's Claude Opus 5.5 placing sixth and Claude Fable ninth. "The world needs a neutral third party to measure how safe and aligned AI actually is once it's in the hands of real people," the company said, a framing that positions the new leaderboard as a hedge against regulatory and enterprise-procurement pressure now building around model behavior, not just raw capability.

Why investors are paying up

The round lands amid growing documentation that AI labs game standard benchmark suites — tuning training data or inference-time behavior to specific test sets rather than improving general capability — which has pushed enterprise buyers toward evaluation providers that run live, continuously refreshed tests instead of static leaderboards. Arena's pitch to investors is that this dynamic favors a company with a large, constantly renewing pool of real user interactions over one relying on a fixed benchmark dataset that degrades in value once it leaks into training corpora. The company says it has no direct competitor running at its scale, though it competes for the same enterprise budgets as human-data and labeling firms such as Scale AI, Surge and Mercor, which have themselves attracted large valuations this year on the premise that model quality increasingly depends on data and evaluation rather than compute alone.

Whether a $3.1 billion valuation for an evaluation layer proves durable will depend on how defensible that data advantage remains as labs build their own internal evaluation pipelines, and on whether enterprise customers keep paying for third-party benchmarking once procurement cycles mature. For now, the round adds Arena to a short list of 2026's AI infrastructure companies — alongside data-labeling and agent-evaluation startups — that have converted benchmark attention into venture capital at a speed that outpaces their revenue growth, a pattern several investors in adjacent rounds this year have described as the market pricing in future necessity rather than current scale.