Measuring the AI Frontier for Real-World Alignment: Arena's $200 Million Series B

These numbers reflect a community, not just a company. Together, we're measuring real-world tasks across agentic capabilities:

• 7 million sessions in Agent Arena in less than 5 months since launch.
• 350 million sessions across the entire Arena platform.
• 62 million votes across text, vision, code, search, video and image modalities.
• Tens of millions of monthly visitors, a global audience from 150+ countries.
• 1,000+ new model evaluations, tracking the frontier across open and proprietary.
• 375k data points open-sourced to the community for open AI research, including our leaderboard methodology.

The scale of that growth is also why we're continuing to expand what Arena measures. “Which model performs best" is no longer the only question that matters. Today, "can we trust what it did and what it says it did" is just as urgent, and very much unanswered. Alongside this raise, we're introducing Arena’s Alignment Index: a first step toward giving the industry the same kind of grounded, real-world verified signal for alignment that we’ve built for capability.

This is the foundation we’ll always ground ourselves in: building and measuring everything against real-world signals. The Alignment Index preview being released today is the next proof point; it starts narrow and is evidence-based by design. We are starting with three signals we can verify against an actual agent trace with real human use: Unauthorized Action (UA), where the model acts beyond what the user asked; False Attribution (FA), where the model attributes a statement, intention, or fact to the user that is contradicted by user-provided evidence; and Deceptive Completion (DC), where the model deceives a user by telling them a task is complete when it is not. These signals are inspired by the same definitions that labs like OpenAI and Anthropic have published in their system cards so that they can serve as an independent assessment of the safety and alignment of AI models. We view our three alignment signals as complementary to our Agent Arena leaderboard, which ranks the capabilities of agents.

We’re sharing the methodology and initial results for 20+ frontier models today. We will continue to update the leaderboard as we expand it signal by signal, the same way our capability leaderboards have grown since day one.

As AI moves faster than most organizations can verify what it's actually doing, we believe this is where an independent, community-driven benchmark needs to go, and we're glad to have the resources to build it properly. Learn more about Arena's Alignment Index methodology and insights in our blog post here.

Arena started as a research experiment and is now becoming how the AI industry checks whether models and agents work and behave safely and truthfully, measured across millions of real interactions. To the community that is helping us keep AI progress focused on real-world benefit: thank you. None of this progress happens without you. We're honored to serve as the voice of humans shaping and improving AI. Let's keep measuring and advancing what the world needs next.

If you're new here, come test with us at:
arena.ai

Love what we do? Join the team at:
arena.ai/company/careers

For any press inquiries contact:
[email protected]