Arena's $3.1bn valuation puts a price on the referee of the AI race

Arena, the AI evaluation company formerly known as LMArena, said on Thursday, October 8, that it had raised a $200m Series B at a valuation of $3.1bn. Lightspeed Venture Partners and Khosla Ventures co-led the round. Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Acrew Capital and Endeavor Catalyst also participated, and existing backers including Andreessen Horowitz, Felicis, AMP PBC, QuantumLight and The House Fund also took part. The company says its annualised revenue now exceeds $100m.

The price has moved quickly. In January Arena raised a $150m Series A at a post-money valuation of $1.7bn, which TechCrunch calculated was nearly doubled by this round in about 10 months. In May 2025 the company's seed round had valued it at $600m. A business that began in 2023 as a university research project ranking chatbots by crowd votes has therefore been repriced roughly fivefold in less than a year and a half. That trajectory says a great deal about what investors now believe the AI economy is short of: not models, but trusted ways of telling them apart.

From student project to toll booth

Arena started as Chatbot Arena, a UC Berkeley project built by Anastasios Angelopoulos and Wei-Lin Chiang and funded initially by grants and donations. Its consumer site is simple. A user types a prompt, two anonymous models answer, and the user votes for the better response. Those votes feed public leaderboards for text, coding, vision, image and video generation, search and, more recently, multi-step agent tasks. Model makers including OpenAI, Google and Anthropic have routinely cited Arena rankings when launching new systems, which turned a free website into something close to an industry scoreboard.

The commercial turn came in September 2025, when the company launched AI Evaluations, a paid service that sells model labs and enterprises performance analytics built on community feedback. By December 2025 it was generating annualised revenue of $30m. On June 29 this year Mr Angelopoulos, the chief executive, wrote that Arena had reached a $100m annualised run rate within eight months of launching the enterprise offering. In the Series B announcement the company says it has now passed that figure.

There is a caveat that investors should keep in view. Mr Angelopoulos told TechCrunch in June, as The Next Web reported, that customers pay on consumption, so the revenue is not recurring in the traditional software sense even though the company describes it as annual recurring revenue. TechCrunch reported in January that Arena itself described the measure as a consumption rate. Usage-based revenue from a small number of very large AI labs can grow fast and fall fast, and it depends heavily on the pace at which those labs release models.

What $3.1bn buys

On the disclosed numbers, the new valuation is roughly 31 times the $100m run rate. That is rich by the standards of software, but not extreme in a market where leading AI application companies have been priced on far higher multiples of current revenue. The more interesting question is what kind of asset investors think they own.

Arena's own pitch, set out in its announcement, is that static benchmarks break down once models recognise they are being tested, and that the industry needs a neutral third party to measure how capable and how safe AI systems really are. The company points to the scale of its data: 350 million sessions across the platform, 62 million votes and more than 1,000 model evaluations, with tens of millions of monthly visitors in more than 150 countries. It also says its Agent Arena product recorded 7 million sessions in under five months.

Alongside the round, Arena released a preview of an Alignment Index that compares 27 models across 90,000 real-world agent sessions. It scores models on three signals drawn from definitions in OpenAI and Anthropic system cards: unauthorised action, where an agent does more than it was asked; false attribution, where it credits the user with statements or intentions that the evidence contradicts; and deceptive completion, where it reports a task as done when it is not. Arena says alignment is improving across model generations, but that frontier systems still fail in consequential ways, such as deleting files without permission or claiming checks they never ran.

This is a deliberate widening of the franchise. A leaderboard of which chatbot writes the better poem is a marketing tool. A measure of whether an autonomous agent can be trusted with a company's files is closer to an audit, and audit functions command durable budgets.

The conflict at the centre of the model

The weakness of the business is the same as its strength. Arena's paying customers include the very labs whose models it ranks, and its value depends on everyone believing that those commercial relationships do not shape the results. That belief has already been tested. In April 2025 a group of researchers, with Sara Hooker as senior author, published a paper called The Leaderboard Illusion. It argued that undisclosed private testing let a few providers try many variants before release and withhold unwanted scores, identified 27 private variants tested by Meta before its Llama 4 launch, and estimated that Google and OpenAI had received about 19.2% and 20.4% of all Arena data, against about 29.7% for 83 open-weight models combined. LMArena strongly denied that its partnerships allowed model makers to game its rankings.

Arena has since published methodology and open-sourced data, saying it has released 375,000 data points including its leaderboard methodology. That is the right response, but the structural tension does not go away as revenue grows. The more of its income comes from a handful of labs, the more each ranking decision looks like a commercial one.

Competitors and substitutes

Crowd-sourced evaluation has not proved an easy market. The Next Web reported in June that Yupp, the only other start-up built on crowd-sourced model comparison, shut down in March after raising $33m. Mr Angelopoulos has said Arena competes for the same budgets as human data and labelling firms such as Mercor, Surge and Scale AI. The labs also run extensive internal evaluations, and an Arena research post in September found that, across 12 models, an AI judge picked its own answer 58% of the time on average while people picked that same answer 34% of the time, an argument for human votes but also a reminder that the field of evaluation is still being invented.

Our view

Arena is one of the more intelligent bets in this cycle because it is a picks and shovels business with a moat made of data and habit rather than compute. Every model launch that cites an Arena ranking reinforces its position, and the move into agent behaviour and alignment points towards a market that will become mandatory once companies hand real work to autonomous systems. But the valuation now prices Arena as infrastructure, and infrastructure for trust is judged by a stricter standard than software. The company's revenue is usage-based and concentrated among the labs it grades, and the Leaderboard Illusion episode showed how quickly questions about independence can surface. For founders, the lesson is that measurement has become a venture category in its own right. For investors, the things to watch are the share of revenue from enterprises rather than model makers, whether the Alignment Index is adopted by buyers and regulators as a reference, and how transparently Arena handles the next dispute over a ranking. Neutrality is the product; it cannot be allowed to become a discount.

Sources

• TechCrunch: Popular AI leaderboard Arena nearly doubles valuation to $3.1B in 10 months
• Arena: Measuring the AI frontier for real-world alignment, Arena's $200 million Series B
• Arena: Arena Alignment Index
• Arena: Arena reaches $100M in 8 months
• The Next Web: Arena AI leaderboard reaches $100 million in revenue
• TechCrunch: LMArena lands $1.7B valuation four months after launching its product
• Reuters via Yahoo Finance: AI startup LMArena triples its valuation to $1.7 billion in latest fundraise
• arXiv: The Leaderboard Illusion
• Arena: Self-preference of AI judges

Filed under Startups. Written by Invest Times Editorial under our editorial standards. Commentary and analysis for information only, not investment advice.