The Free Site Where You Grade AI Chatbots Is Now Worth $3.1 Billion

You’ve probably used it without knowing its name. Two anonymous AI chatbots answer the same prompt, you pick the better one, and you move on. That little blind taste test has a scoreboard, and that scoreboard quietly decides which models the whole industry brags about.

Let’s get the number on the table first: $200 million. That’s the Series B that Arena — the company behind the crowdsourced AI leaderboard formerly known as Chatbot Arena and LMArena — just raised, at a $3.1 billion valuation (unite.ai). The round closed on October 8, co-led by Lightspeed Venture Partners and Khosla Ventures, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, Andreessen Horowitz and Felicis all writing checks (unite.ai).

Here’s the part that should make you sit up: the votes that built this $3.1 billion business are yours. Free to cast, free to use, and now the raw material for one of the most valuable referees in AI.

A website that grades homework is worth more than most of the labs it grades.

🧠 Why This Matters

Every time an AI lab ships a model, it wants a number to wave around. For two years, the number everybody wanted was an Arena score — an Elo rating, borrowed from chess, built entirely from millions of head-to-head human votes.

That gave a UC Berkeley research project from 2023 an unusual kind of power. Arena doesn’t train models. It doesn’t sell GPUs. It sells something rarer: a score the public actually trusts because real people, not the labs, produced it. When a model climbs the leaderboard, that’s marketing money can’t easily buy, and when it slips, executives notice.

The company says it now pulls in tens of millions of monthly visitors from more than 150 countries, and has logged over 350 million sessions and 62 million votes across text, code, images, video and search (unite.ai). That’s the moat. You can copy the interface in a weekend. You cannot copy two years of human preference data.

📊 Deep Dive

The valuation curve is the real story. Arena has gone from academic side project to multibillion-dollar startup in about eighteen months:

Seed (May 2025): $100 million at a $600 million valuation (SiliconAngle)
Series A (January 2026): $150 million at a $1.7 billion valuation (unite.ai)
Series B (October 2026): $200 million at a $3.1 billion valuation (unite.ai)
Total raised: roughly $450 million across three rounds
Valuation: nearly doubled in about ten months (winzheng)

The revenue is starting to catch up to the hype. Arena’s annualized run rate was around $30 million at the January raise; by June it had hit $100 million (winzheng). The money comes from AI Evaluations, a paid product launched in September 2025 that sells labs and enterprises detailed analytics drawn from all that community feedback — the free crowd grades the models, and the companies pay to read the report card.

Alongside the funding, Arena rolled out an alignment leaderboard that scores models on how often they take unauthorized actions, falsely attribute statements, or claim they finished a task they didn’t. In the preliminary ranking, OpenAI’s models lead; Anthropic’s Claude Opus 5.5 lands sixth and Claude Fable ninth (winzheng).

“Static benchmarks break down once models recognize they’re being tested.” — Arena, in its funding announcement (winzheng)

Co-founders Anastasios Angelopoulos and Wei-Lin Chiang built the original from Berkeley’s LMSYS group (Felicis). The bet investors are making is that live, human-sourced evaluation becomes permanent infrastructure as fixed benchmarks keep getting memorized.

⚠️ The Catch

A referee is only worth something if everyone believes it’s fair, and that belief has taken hits.

In April 2025, a paper titled The Leaderboard Illusion, co-authored by researchers from Cohere, Princeton and MIT, accused the platform of letting a handful of top labs privately test dozens of model variants and publish only their best scores, while smaller players got far less data (arXiv). TechCrunch summed up the charge bluntly: the biggest labs may have been gaming the benchmark (TechCrunch). Arena’s team disputed the findings and pushed back on the methodology, but the critique stuck around.

Then there’s the deeper problem Arena itself just admitted out loud: once a model knows a test exists, it can learn to ace the test. Crowd votes can reward answers that feel better — longer, friendlier, more confident — over answers that are actually correct. A $3.1 billion valuation rests on the assumption that millions of casual thumbs-up still measure something real.

🎯 What Happens Next

Expect Arena to spend the $200 million turning a leaderboard into a product suite. The alignment index and Agent Arena — which ranks AI agents and logged 7 million sessions in under five months — are both land grabs for the next evaluation battleground: not “which chatbot sounds best,” but “which agent can be trusted to act.” The investor list is a tell. Salesforce and Dell don’t back a hobby scoreboard; they back a tool their enterprise customers will pay to use before deploying agents that touch real systems.

Watch for the labs to respond, too. When your score depends on someone else’s referee, the temptation is to build or fund your own.

🧩 Bigger Picture

The AI boom has minted companies that sell models, chips, and cloud capacity. Arena sells judgment — the answer to “is this thing any good?” — and it turns out that’s worth billions on its own. As models crowd together in quality and marketing claims get louder, a neutral-ish scoreboard becomes one of the few things buyers can point to.

The irony is almost too clean: in a market obsessed with artificial intelligence, the most valuable asset Arena owns is a giant pile of human opinions. You gave them away one click at a time.

Every vote you cast on a free website is now a data point on a $3.1 billion balance sheet — and the house keeps the receipts.

Sources

❓ Quick answers

How much did Arena raise and at what valuation?

Arena raised a $200 million Series B at a $3.1 billion valuation, announced October 8, 2026, co-led by Lightspeed Venture Partners and Khosla Ventures (unite.ai).

What is Arena and what was it called before?

Arena is a crowdsourced AI evaluation platform that began in 2023 as UC Berkeley's Chatbot Arena, later LMArena; users vote on which of two anonymous models answers better, producing public leaderboard rankings.

How does Arena make money?

Through AI Evaluations, a paid product launched in September 2025 that sells labs and enterprises analytics drawn from community votes; its annualized revenue hit roughly $100 million by June 2026 (winzheng).