Arena raises $200M at $3.1B valuation as AI evaluation becomes a business

Arena lets people send the same prompt to two unnamed AI models, compare the responses and vote for the better one. The company turns those interactions into public leaderboards and paid evaluation reports for AI labs and enterprises. Arena said Thursday it raised a $200 million Series B at a $3.1 billion valuation.

The workflow now reaches beyond single answers. In Agent Arena, the platform follows each turn of a longer task—such as coding, document analysis or creative work—and uses the trace to measure whether an agent completed what the user asked. Those signals give model builders a view of performance on real work that fixed benchmark questions can miss.

The public leaderboard is free; the commercial product sells deeper analytics built from community comparisons. TechCrunch reported in June that Arena charges customers based on consumption rather than a recurring subscription, even though the company describes its metric as annualized revenue. Arena now says that figure exceeds $100 million. It is a fast revenue run rate, but not the same thing as $100 million of contracted recurring revenue.

The financing arrives with a preview of Arena’s Alignment Index. It checks agent traces for three behaviors: acting outside a user’s instructions, falsely attributing a statement or intention to the user, and claiming a task is finished when it is not. That moves Arena from ranking which response people prefer toward testing what autonomous software actually does during multi-step work.

Lightspeed Venture Partners and Khosla Ventures co-led the round. Arena raised a $150 million Series A in January at a $1.7 billion post-money valuation, TechCrunch reported. The new $3.1 billion mark is $1.4 billion higher nine months later, although Arena’s announcement did not specify whether the latest valuation is pre- or post-money.

That rapid commercialization makes Arena’s independence more valuable and harder to demonstrate. Salesforce Ventures, which joined the round, wrote that the leaderboard depends on no model lab owning the scoreboard and noted that Arena publishes its methods and opens data to scrutiny. Arena sells evaluations to the same class of AI labs whose models populate its rankings; trust in the methodology is therefore part of what paying customers are buying.