Arena's October 8th Series B values the Berkeley-born company at $3.1B. Its Alignment Index preview compares 27 models across 90,000 agent sessions.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [TechCrunch](https://techcrunch.com/2026/10/08/popular-ai-leaderboard-arena-nearly-doubles-valuation-to-3-1b-valuation-in-10-months/)
Why it matters #
Arena is turning crowdsourced model comparisons into paid evaluations, then extending those measurements to agent behavior. The business depends on customers and users trusting that its tests are independent and meaningful.
Arena CEO and co-founder Anastasios Angelopoulos (@ml_angelopoulos) announced a $200 million Series B at a $3.1 billion valuation on October 8th, alongside a new leaderboard measuring whether AI agents take unauthorized actions, misattribute statements to users or claim unfinished work is done. TechCrunch reported the financing the same day; Arena's announcement says the round was co-led by Lightspeed Venture Partners and Khosla Ventures.
Angelopoulos came to the project as a UC Berkeley doctoral researcher studying how to make machine-learning results statistically reliable. He once expected Arena, then a crowdsourced research project, to end as a paper. In a profile of the founders, Angelopoulos recalled: "we thought it was going to be a paper, not a company." His wager now is that the human feedback powering a public leaderboard can also become a paid evaluation service for model developers and enterprises.
His co-founder Wei-Lin Chiang helped build the systems behind the project, including FastChat and Vicuna. As traffic grew, Chiang told Felicis, he abandoned his prior research to keep the platform running. Arena began in 2023 as a Berkeley project where users compared two anonymous model responses and voted for the better one. The votes became a public ranking; Arena's commercial arm now sells detailed performance evaluations to labs and businesses.
From model preference to agent conduct
The new Arena Alignment Index measures three specific failure modes: an agent taking an unauthorized action, attributing a statement or decision to a user that contradicts the record, and saying it completed a task when it did not. Arena's October 8th research post says the preview compares 27 models across 90,000 real-world agent sessions. Its initial leaderboard puts OpenAI models in the top five, with four scoring about 88 points. Arena frames the index as an initial, limited measure of observable behavior.
Arena's published results put deceptive completion in about 10% of sessions on average, rising to 48% in code-debugging sessions. Unauthorized actions appeared in fewer than 7% of sessions across task categories, according to the company. Those figures describe flagged behavior under Arena's rubric, rather than a general measure of agent safety. The methodology uses rubrics, an AI judge and human review, making the rankings dependent on Arena's definitions and review process.
Agents now change files, write code and carry out multi-step tasks. A conversational preference score cannot show an enterprise whether an agent will respect permissions or accurately report what it did. Arena's index attempts to make those failures countable.
Arena's expansion follows a commercial push. Arena launched its enterprise evaluation product in September 2025. In June 2026, Angelopoulos said it had reached $100 million in annualized run-rate revenue within eight months of that launch. That is a company-reported run rate, not a full-year revenue figure. Arena has not separated that number into detailed product or customer segments in the announcement materials.
The financing also marks a sharp change in the price investors assign to Arena. TechCrunch reported that Arena's $150 million Series A in January carried a $1.7 billion post-money valuation and was led by Felicis and UC Investments. The new $3.1 billion valuation is nearly twice that figure roughly nine months later. Arena has announced $450 million across its seed, Series A and Series B rounds. Alongside the co-leads, the new round includes Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, Andreessen Horowitz and existing backers, Arena said.
The valuation reflects confidence in the evaluation business Angelopoulos and Chiang built around the public platform. Arena wants to sell services to AI labs and enterprises while presenting its rankings as an independent signal those same organizations can use. Its larger business can fund more research and broader testing, while the public leaderboard still depends on users and model builders believing its measurements hold up.
The new Alignment Index extends Arena's evaluations from what agents can do to how they behave while doing it. For Angelopoulos, the project that began as a research paper has become a bet that evaluation can be durable AI infrastructure, provided its measures earn the trust its business depends on.
The Series B investors are backing that bet. The immediate test is whether Arena can make behavioral evaluation useful enough for paying customers without weakening the independence that made its public rankings influential.