A platform that began as a university research experiment is now valued at $3.1 billion. Arena, which launched in 2023 out of UC Berkeley as a crowd-powered ranking system for artificial intelligence models, announced Thursday that it has closed a $200 million Series B funding round at that valuation.

The announcement follows a milestone the company disclosed in June, when it reported reaching $100 million in annualized run-rate revenue - a figure that underscores how quickly the business has scaled since transitioning from academic project to commercial enterprise.

The Series B was co-led by Lightspeed Venture Partners and Khosla Ventures. A broad group of additional investors participated in the round, including Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, Andreessen Horowitz (a16z), Felicis, and several others. The company's previous fundraise, a $150 million Series A announced in January of this year, valued Arena at $1.7 billion post-money. At that point, the company was generating roughly $30 million in annualized revenue. The near-doubling of its valuation in approximately ten months reflects both accelerating revenue growth and rising investor confidence in the AI evaluation space.

How the Platform Works

Arena's core consumer product is a free, publicly accessible platform where users submit prompts or request AI-generated projects and then compare the outputs of competing models side by side, voting on which response they find more useful or accurate. The company says the platform draws tens of millions of visitors each month, generating a continuous stream of real-world human preference data that forms the backbone of its evaluation methodology.

That community-driven approach sets Arena apart from conventional AI benchmarks, which typically rely on fixed datasets and automated scoring. Because Arena's ratings come from actual users interacting with models in open-ended ways, the results are considerably harder for AI systems to game through targeted training.

A Commercial Product Launched at the Right Moment

In September 2024, Arena introduced its paid commercial offering, branded as AI Evaluations. The service gives AI model developers and enterprise customers access to in-depth performance analytics derived from the platform's community feedback. The timing turned out to be particularly well-suited to market conditions.

Throughout this year, growing scrutiny has been directed at the AI industry over the reliability of standard benchmarks. Researchers and practitioners have observed that leading AI models appear to be optimized - whether deliberately or as an emergent effect of training - to perform well on known benchmarks without necessarily demonstrating equivalent capability in real-world scenarios. Simultaneously, businesses deploying AI internally have grown skeptical of relying solely on published leaderboard scores, preferring evaluations that reflect their specific use cases and user bases.

Arena's product addresses both concerns. For model developers, it offers a credible, third-party signal of how their systems perform when tested by genuine users rather than curated test sets. For enterprises, it provides a framework for identifying which models best serve their particular operational needs.

"AI is advancing faster than our ability to evaluate it, and static benchmarks break down once models recognize they're being tested. The world needs a neutral third party to measure how safe and aligned AI actually is once it's in the hands of real people. Arena is stepping into that role today."

A New Leaderboard Category: Alignment

Alongside the funding announcement, Arena revealed an expansion of its leaderboard to include a dedicated alignment category. This new ranking evaluates AI models not just on the quality or accuracy of their outputs, but on behavioral characteristics that bear directly on safety and trustworthiness.

The alignment leaderboard tracks several specific failure modes that have become points of concern as AI systems grow more capable and are deployed in higher-stakes contexts. These include:

• Unauthorized action - instances where a model takes steps or performs actions beyond what the user explicitly requested.
• False attribution - cases in which a model incorrectly credits statements, quotes, or facts to the wrong source.
• Deceptive completion - situations where a model falsely indicates that it has completed a task when it has not actually done so.

The leaderboard is currently in a preliminary state, but early results show a cluster of models from OpenAI occupying the top positions. Anthropic's Claude Opus 5.5 ranks sixth, while Claude Fable appears in ninth place on the current standings.

The introduction of an alignment-focused ranking reflects a broader shift in how the AI industry is approaching model evaluation. As capabilities advance, the question of whether a model behaves reliably, honestly, and within intended boundaries has become as important to many stakeholders as raw performance metrics. Arena's expansion into this area positions the company at the center of what is likely to be an increasingly consequential debate about how AI systems should be measured and held accountable.