Arena, the company behind the popular LMArena leaderboard

has raised $200 million at a $3.1 billion valuation. The round, announced Thursday, was led by Lightspeed Venture Partners and Khosla Ventures, with participation from Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, a16z, Felicis and others. Only in January the company closed a $150 million Series A at a $1.7 billion post-money valuation — meaning the valuation has nearly doubled in roughly ten months. In June Arena disclosed $100 million in annual recurring revenue, up from $30 million at the time of the previous round.

From lab to commercial product

The platform began in 2023 as a research project at UC Berkeley, ranking models through crowdsourcing: users feed prompts or request "vibe coding" projects and vote on which model performed better. The company says the site draws tens of millions of monthly visitors and remains free for consumers. Last September Arena launched its first commercial product, AI Evaluations, which sells detailed performance analytics based on that community feedback to model labs and enterprises.

Why investors moved now

The timing proved precise. This year AI labs realized their models were "gaming" static benchmarks — finding ways to score highly without actually earning it. At the same time, enterprises were looking for help picking the right model for internal needs instead of relying on standard benchmarks alone. "AI is advancing faster than our ability to evaluate it, and static benchmarks collapse the moment models realize they're being tested," the company said in its fundraising announcement. "The world needs a neutral third party to measure how safe and aligned AI is when it's in the hands of real people. Arena is stepping into that role today."

A new alignment category

Accordingly, Arena has added a preliminary "alignment" category to the leaderboard. It ranks models on issues such as unauthorized action — performing actions the model was not asked to perform; false attribution — attributing statements or facts to the wrong source; and what the company calls "deceptive completion" — lying about completing tasks that were not actually performed. For now, a slate of OpenAI models leads the preliminary alignment table, with Anthropic's Claude Opus 5.5 and Claude Fable sitting in sixth and ninth place respectively.