HomeAIAI Funding & DealsArena Raises $200M at $3.1bn Valuation to Rank AI Models

Arena Raises $200M at $3.1bn Valuation to Rank AI Models

Arena, the public leaderboard where people vote on which AI model gives the better answer, has raised $200 million in a Series B round at a $3.1 billion valuation. The company announced the round on 8 October, saying it has passed $100 million in annualised revenue.

Lightspeed Venture Partners and Khosla Ventures co-led the round, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Acrew Capital and Endeavor Catalyst joining. Existing backers including a16z, Felicis, AMP PBC, QuantumLight, The House Fund and Panos Madamopoulos also took part, according to Arena. Salesforce Ventures, one of the new investors, set out its reasons in an investment note, arguing that fixed benchmarks “saturate quickly” and that their test questions often leak into training data.

From Berkeley project to $3.1bn in under three years

Arena started in 2023 as a research project at UC Berkeley, TechCrunch reported. In January it raised a $150 million Series A at a $1.7 billion post-money valuation, when its annualised revenue was about $30 million. TechCrunch says it reached a $100 million run rate in June, and that the valuation has nearly doubled in about ten months.

The method dates from that research phase. The founding team, including co-founders Wei-Lin Chiang and Ion Stoica, described it in a 2024 paper on Chatbot Arena: users put one prompt to two anonymous models, pick the better answer, and the votes are turned into a ranking. That design is why a free consumer site became a measuring tool that model makers watch closely.

The consumer site stays free. The money comes from AI Evaluations, a commercial product launched in September 2025. The revenue numbers are Arena’s own and have not been independently audited. Arena’s round joins a busy run of AI raises tracked on TSN’s AI funding rounds hub. Company-stated run rates, like the one ElevenLabs gave recently, have become a common yardstick in these deals, but they are not the same as published accounts.

Scale, by Arena’s count

Arena’s announcement gives a set of usage figures: 350 million sessions, 62 million votes, tens of millions of monthly visitors from more than 150 countries, and over 1,000 new model evaluations. It says it has open-sourced 375,000 data points and that its Agent Arena logged seven million sessions in under five months. In a post on X the company said it is “still a nimble team of 90” and will spend the money on “safety signals, agentic capabilities, and modalities”.

Arena’s pitch is that fixed tests are losing their value. “AI is advancing faster than our ability to evaluate it, and static benchmarks break down once models recognize they’re being tested,” the company wrote. “The world needs a neutral third party.”

Measuring behaviour, not just answers

Alongside the raise, Arena previewed an Alignment Index. It scores more than 20 frontier models on three signals of agent misbehaviour, which it calls Unauthorized Action, False Attribution and Deceptive Completion. Arena says the definitions draw on safety terms in OpenAI’s and Anthropic’s model system cards. TSN covered the index and its early scores in more detail in two new ways of checking whether AI agents overstep or fib.

Arena’s paid evaluations put its neutrality under scrutiny

Arena’s value rests on being seen as neutral, and the funding sharpens that question rather than settling it. Model makers, whose systems compete on the leaderboard, are an obvious market for paid evaluations. That kind of overlap is not unusual. Credit rating agencies are paid by issuers, and audit firms by the companies they audit. But those industries learned, at some cost, that the arrangement only holds up when methods are published, conflicts are disclosed and the same rules apply to every client.

The Alignment Index raises the stakes. A ranking of which chatbot writes better poetry is a matter of taste. A score for whether an agent takes actions it was not asked to take starts to look like a safety rating, and companies may cite it when buying agent software. Compliance and risk teams would be wise to treat any single score as one input, ask how it was produced and check whether the vendor being rated is also a paying client. Arena’s methods have been challenged before. In “The Leaderboard Illusion”, an April 2025 paper, outside researchers said undisclosed private testing let a handful of providers try many model variants before release and publish only the best score, and estimated that Google and OpenAI had received about 19.2% and 20.4% of all Arena data respectively. Arena says it has since open-sourced 375,000 data points, including its leaderboard methodology. Its investors make neutrality the selling point too: Salesforce Ventures writes that the scoreboard “only works if no lab owns” it. Independent testing of agent behaviour is clearly needed: the gaps METR and Redwood found in OpenAI’s agent monitoring show how much can be missed from inside.

With $200 million in the bank and a team of 90, Arena can afford to publish its evaluation methods and its client conflicts in full. Whether it does will say more about its neutrality than any leaderboard.

Sources

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.