Arena, the website where people vote on which AI model gives the better answer, raised $200 million at a $3.1 billion valuation on October 8, 2026. Lightspeed Venture Partners and Khosla Ventures co-led the Series B, TechCrunch reports. The price nearly doubles what investors paid about ten months earlier. With the money, Arena is pushing beyond "which model is smartest" toward rating how safely models behave.
The round
Arena announced the deal on its own blog. Other new investors include Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst. Existing backers a16z, Felicis, AMP PBC, QuantumLight and The House Fund also joined, according to Arena.
The company's value has climbed fast:
| Round | Date | Amount | Valuation |
|---|---|---|---|
| Series A | January 6, 2026 | $150 million | $1.7 billion post-money |
| Series B | October 8, 2026 | $200 million | $3.1 billion |
Felicis and UC Investments led the Series A, Arena's blog says. TechCrunch reported the $1.7 billion valuation in June. Arena did not say how it will spend the new money, beyond building its products and hiring.
How Arena makes money
Arena began in 2023 as a research project at UC Berkeley that crowdsourced AI model rankings. It became a company in April 2025, TechCrunch reported in June. Its co-founders are chief executive Anastasios Angelopoulos and chief technology officer Wei-Lin Chiang. UC Berkeley professor Ion Stoica, a co-founder of Databricks, advised the project before it incorporated.
The public leaderboard is free. A user types a prompt, two models answer, and the user picks the better reply. Arena says the platform has logged 350 million sessions and 62 million votes, from people in more than 150 countries.
Money comes from AI Evaluations, a paid service launched in September 2025 that gives model makers and companies detailed analysis. Arena now says it has passed $100 million in annualized revenue. That is up from $30 million at the January round, TechCrunch reported. Angelopoulos has said Arena charges for "consumption," so the figure is usage-based rather than recurring subscription revenue.
Arena told TechCrunch in June that it has no direct competitors. It says it competes "for the same dollar" with data firms such as Mercor, Surge and Scale AI.
A new index for unsafe behavior
Alongside the round, Arena launched a preview of the Arena Alignment Index. It scores models on three signals:
- Unauthorized Action
- False Attribution
- Deceptive Completion
Going by the names, they cover a model acting without permission, crediting the wrong source and claiming work it did not finish. Arena says the definitions draw on system cards from OpenAI and Anthropic, the documents labs publish about a model's risks. In the preliminary results, OpenAI models sit at the top, TechCrunch reports. Claude Opus 5.5 ranks sixth and Claude Fable ninth.
Arena argues the field needs this kind of testing. "AI is advancing faster than our ability to evaluate it," the company wrote. It wants to be "a neutral third party to measure how safe and aligned AI actually is."
What this means for developers
- Use the leaderboard as one signal, not the answer. Arena ranks models on what crowds prefer. Your own tasks, data and costs still decide which model fits.
- Watch the Alignment Index for agent work. If your product lets a model act on its own, unauthorized actions and false "done" messages are the failures that hurt. A public score for them is useful, even in preview.
- Expect labs to chase the new index. A public safety ranking gives model makers a new number to improve. Check whether those gains hold up in your own tests.
- Know who pays Arena. Its paid evaluation service is sold to AI model labs and companies. Keep that in mind when you read the rankings.
The next sign of whether the index matters is whether labs start citing it in their own model launches.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Top comments (0)