Arena

Arena is launching the Alignment Index, a public leaderboard that scores how frontier AI models behave when real people use them to get work done. The index evaluates 27 models across more than 90,000 real-world agent sessions , grading them on three concrete failure modes: Unauthorized Action, where the model acts beyond what the user asked, False Attribution, where the model attributes a statement or fact to the user that is contradicted by their own evidence, and Deceptive Completion, where the model tells a user a task is done when it is not. These signals are inspired by the same definitions that labs like OpenAI and Anthropic have published in their system cards , which lets the index function as an outside check on claims the labs make about their own models. The launch is paired with a $200M Series B at a $3.1B valuation, co-led by Lightspeed Venture Partners and Khosla Ventures, with participation from Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Acrew Capital, Endeavor Catalyst, and existing investors including a16z, Felicis, AMP PBC, QuantumLight, and The House Fund . The company is the same group behind Chatbot Arena and the Agent Arena leaderboard, started as a research project out of UC Berkeley's Sky Computing Lab by Wei-Lin Chiang, Anastasios N. Angelopoulos, and Ion Stoica . The Alignment Index sits alongside their existing capability rankings, extending the methodology from "which model is best" to "which model can be trusted to do what it was asked." The timing matters for anyone shipping agents into production. Arena is releasing methodology and initial results for 20+ frontier models and plans to expand the benchmark signal by signal, giving builders, procurement teams, and policy watchers a third-party reference point as agents move from answering questions to taking actions on users' behalf in code, analysis, and document work.