Arena, valued at $3.1bn, wants to grade AI safety

Arena, the company behind the crowdsourced AI model leaderboard formerly known as LMArena, has raised $200m in a Series B round at a $3.1bn valuation and plans to rate models on safety as well as performance, the company said on Thursday. Bloomberg reported the same day that the round values the business, which began as a University of California, Berkeley research project, at over $3bn. Arena's chief executive and co-founder is Anastasios Angelopoulos, according to the company.

According to Arena's announcement, the round was co-led by Lightspeed Venture Partners and Khosla Ventures, with participation from Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Acrew Capital and Endeavor Catalyst. Existing investors named include a16z and Felicis. TechCrunch's report lists the same lead investors but does not name Acrew. The announcement does not say whether $3.1bn is a pre-money or post-money figure; TechCrunch described the January valuation of $1.7bn as post-money.

Arena announced a $150m Series A on 6 January, led by Felicis and UC Investments. By our arithmetic, $3.1bn against $1.7bn is a rise of about 82 per cent in nine months, and the two rounds together come to $350m. On revenue, Arena said it has "exceeded $100M in annualized revenue"; in June it said it had passed a $100m run-rate within eight months of launching its enterprise offering. TechCrunch reported that the figure was $30m at the time of the Series A. By our arithmetic the new valuation is about 31 times the $100m run-rate. The numbers are the company's own, and an annualised run-rate is not recognised revenue.

The business model sits behind the safety pitch. The platform is free for consumers, who enter prompts and vote on which model answers better, TechCrunch reported. Arena's commercial product, AI Evaluations, launched in September 2025 and gives model labs and enterprises analytics drawn from that community feedback, it said. Model developers are therefore customers of the paid service as well as the subjects of the public rankings. For its agent leaderboard, Arena says it uses a "causal inference methodology" that observes the full human-agent workflow.

The new Alignment Index is the safety product. Arena says it starts with three signals that can be checked against an agent's actual trace: Unauthorized Action, where a model acts beyond what the user asked; False Attribution, where it attributes to the user a statement, intention or fact contradicted by the user's own evidence; and Deceptive Completion, where it tells a user a task is complete when it is not. The company says the index compares 27 models across 90,000 real-world agent sessions. It calls the release a preview that "starts narrow", and says the signals are "inspired by the same definitions that labs like OpenAI and Anthropic have published in their system cards".

Arena framed the move as a gap in the market. "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 in its announcement. OpenAI models hold the top places on the preliminary alignment leaderboard, TechCrunch reported, with Anthropic's Claude Opus 5.5 in sixth place and Claude Fable in ninth. Neither the announcement nor TechCrunch's report carries a comment from a named executive or investor. We have not independently reviewed Arena's methodology or scores.

Arena's own usage figures do not line up across its posts. The June announcement cited more than 82m votes and more than 700m conversations; the Series B post cites 62m votes and 350m sessions across the platform. By our subtraction, the vote count is 20m lower in the later post. The two posts may count different things, and the later one describes its votes as spanning text, vision, code, search, video and image, but neither explains the difference.

Arena's neutrality has been questioned before. A paper titled The Leaderboard Illusion, first posted in April 2025 by Shivalika Singh and colleagues, alleged that undisclosed private testing let a handful of providers test multiple model variants and withhold unwanted scores, and that proprietary models received a disproportionate share of the platform's data. The paper concerns the performance rankings, not the Alignment Index, and we did not review Arena's response to it.

The question the round raises is who gets paid to judge the models. Arena sells to developers whose models it also ranks, while its safety pitch rests on being a "neutral third party". Its answer is method: community voting, evidence taken from agent traces, and open data, with 375,000 data points released, it said. The index covers three behaviours for now. Who should grade frontier models is an open question elsewhere too: the largest labs are working on a safety standards body of their own, and Komposite reported on Friday on TypeSafe AI's $870m raise.