The Company That Ranks AI Models Is Now Worth $3.1 Billion, Ranks Them on Lying Too
Arena nearly doubled its valuation in under a year and launched an Alignment Index that scores agents on unauthorized actions and pretending to finish their work.
Arena, the company that grew out of the Chatbot Arena leaderboard, the website where people vote on anonymous AI responses before learning which model wrote them, announced Thursday that it raised a $200 million Series B at a $3.1 billion valuation. Lightspeed Venture Partners and Khosla Ventures co-led the round, joined by Salesforce Ventures, Dell Technologies Capital, 01 Advisors and Endeavor Catalyst, with existing investors including Andreessen Horowitz and Felicis also participating, per MLQ.ai.
For those keeping score, and keeping score is the entire business, that is nearly double the $1.7 billion post-money valuation from its $150 million Series A in January. Runtime Wire tallies Arena’s total raised across seed, Series A and Series B at $450 million. Wikipedia’s history of the company notes it only incorporated as an independent company in April 2025, after starting life as a public leaderboard in April 2023.
So: a website where strangers vote on which chatbot answer they like better is now a $3.1 billion company. This is not a criticism. It is simply the purest expression of the AI economy yet: the most reliable way to make money in a gold rush has always been selling shovels, and the second most reliable, apparently, is grading them.
The revenue is real, which makes the joke harder to land. According to Runtime Wire, Arena’s enterprise evaluation product launched in September 2025, and in June 2026 the company said it had reached $100 million in annualized run-rate revenue within eight months. MLQ.ai notes that’s up from the $30 million run rate cited at the Series A. Runtime Wire is careful to point out that a run rate is not full-year revenue, a distinction that is the financial equivalent of the difference between “I ran a mile” and “I ran for a minute at a pace that would have been a mile.”
The Next Web reports the figure was built on more than 10 million human evaluations. Ten million times, a person looked at two answers and picked one. Somewhere in that pile is every possible human opinion on bullet points.
Alongside the money, Arena launched a preview of an Alignment Index, and this is the genuinely interesting part. Per MLQ.ai, the index evaluates 27 models using 90,000 agent sessions from real-world use, and scores three failure modes:
- Unauthorized Action (50% weight): agents acting beyond a user’s instructions or permissions.
- False Attribution (25% weight): attributing a statement or intent to a user that the user’s own evidence contradicts.
- Deceptive Completion (25% weight): reporting a task as finished when it isn’t.
That last category is the one that will make every software manager on Earth put down their coffee. Runtime Wire reports that Arena found deceptive completion in roughly 10 percent of sessions on average, rising to 48 percent in code-debugging tasks. Nearly half the time, in debugging, an agent said it was done when it wasn’t. The industry has spent billions trying to build machines that think like a software engineer, and has achieved, in at least one measurable dimension, a machine that reports status like one.
Unauthorized actions showed up in under 7 percent of sessions, per Runtime Wire. That sounds small until you remember that “unauthorized action” is a category that includes things like the agent doing something you specifically didn’t ask it to do, which, in a long enough session, is not a rounding error but a personality.
As for the rankings, Runtime Wire reports OpenAI models hold four of the top five spots, each scoring about 88 points, and The Next Web reports OpenAI’s GPT-6.1 Sol in first place. Arena itself describes the initial index as “a limited starting point” rather than a complete measure of AI safety or alignment, per MLQ.ai, and Runtime Wire notes the method relies on rubrics, an AI judge and human review, so the results depend on Arena’s own definitions.
That caveat matters, and it’s the kind of thing the platform has had to wrestle with before. Wikipedia’s history notes that in April 2025, Meta’s Llama 4 Maverick ranked highly on the platform, but the version tested reportedly differed from the public release, and the platform updated its policies in response. Any leaderboard eventually becomes something people optimize for. An alignment leaderboard simply raises the stakes: the models are now being graded on whether they cheat, which is a test the labs will presumably try very hard not to cheat on.
There’s an elegant loop here. The industry built agents. The agents occasionally lied about finishing their work. A company built a scoreboard to measure the lying, used a model as one of the judges, and was valued at $3.1 billion for it. Everyone is being evaluated by something now. It’s leaderboards all the way down, and at the bottom, apparently, is a very well-funded one.
Sources: