Arena Raises $200 Million at $3.1 Billion and Sells Evaluations to the Labs It Ranks

Arena, the AI model leaderboard that began as a UC Berkeley research project, announced a $200 million Series B on Thursday at a $3.1 billion valuation, according to its own announcement. Lightspeed Venture Partners and Khosla Ventures co-led the round. Neither Arena nor TechCrunch says whether $3.1 billion is a pre-money or post-money figure, and that gap changes what the number means.

The round is announced, not independently confirmed by a filing, and it arrives a day after Mecka AI's $60 million raise, another startup whose product is data sold to AI developers. Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst took part, and Arena lists a16z, Felicis, AMP PBC, QuantumLight and The House Fund among existing investors.

From $1.7 billion to $3.1 billion in 10 months

Arena, formerly LMArena, raised a $150 million Series A in January 2026, led by Felicis and UC Investments, per its own post. TechCrunch puts that round at a $1.7 billion post-money valuation, so the new figure is an 82 percent step up.

Round	Date	Amount	Valuation
Series A	Jan. 2026	$150 million	$1.7 billion post-money
Series B	Oct. 8, 2026	$200 million	$3.1 billion, basis unstated

1. Series A, Jan. 20261.7 $ billion
2. Series B, Oct. 20263.1 $ billion

Source: TechCrunch and Arena announcements, accessed 2026-10-08

Revenue explains part of the move. TechCrunch reports annualised revenue of $30 million at the Series A and $100 million by June 2026, and Arena says it has since exceeded $100 million. A run rate multiplies a recent period forward, so it is not recognised revenue. On those figures the Series A priced Arena at about 57 times run-rate revenue, and the Series B at no more than 31 times. If $3.1 billion is post-money, the pre-money value was $2.9 billion.

What Arena sells, and to whom

Arena runs a free consumer platform that it says draws tens of millions of monthly visitors. The revenue comes from AI Evaluations, launched in September 2025, which gives model labs and enterprises performance analytics built on community feedback, per TechCrunch.

That makes the labs both the subjects of the rankings and the buyers of the analytics. Neither TechCrunch's report nor Arena's post as retrieved addresses how the two businesses are kept apart. Arena's post calls the company "the most trusted evaluation platform in AI". Trust is the asset being priced, and checking AI claims is contested elsewhere too: OpenAI withdrew three math papers a day after release.

Arena's own scale figures are 62 million votes, 350 million sessions and more than 1,000 model evaluations. Its Agent Arena, which scores models doing multi-step tasks, logged 7 million sessions in under five months.

The Alignment Index, and a ranking conflict

The announcement introduced the Arena Alignment Index, which scores models on three behaviours, each checked against agent traces from real human use.

Signal	What it flags
Unauthorized Action	The model acts beyond what the user asked
False Attribution	The model credits the user with a statement the evidence contradicts
Deceptive Completion	The model says a task is done when it is not

TechCrunch reports that OpenAI models top the preliminary alignment leaderboard, with Claude Opus 5.5 sixth and Claude Fable ninth. A related Arena article, covering 27 models across 90,000 agent sessions, names GPT-6.1 Sol, Claude Opus 5.5 and Grok 4.7 as top scorers. Those two accounts place Opus 5.5 differently, and the live leaderboard at arena.ai was not checked against either. The index counts three signals; Arena says it will add more over time.

What would confirm the valuation

The test is whether the evaluations revenue keeps compounding after Arena publishes rankings that some paying customers will dislike. Arena has said it will add signals to the index over time, and each addition reshuffles a table its customers read closely.