Arena is introducing an Alignment Index to evaluate AI agents in real-world sessions, focusing on three signals—Unauthorized Action (50%), False Attribution (25%), and Deceptive Completion (25%)—to gauge alignment based on crowdsourced user feedback and real-world traces.

The index assesses how often AI models take unauthorized actions or falsely claim completion, using crowdsourced judgments to rate performance across 27 models over 90,000 real-world agent sessions, guided by rubrics refined with human input.

Inspiration for the framework comes from OpenAI and Anthropic system cards, with an emphasis on how deviations from human values manifest in long, real-user interactions.

Arena closed a $200 million Series B at a $3.1 billion valuation, in a round co-led by Lightspeed Venture Partners and Khosla Ventures, with participation from Salesforce Ventures and a lineup of strategic backers.

The company is rapidly expanding revenue, reporting more than $100 million in annualized revenue only eight months after launching its enterprise offering.

This Series B follows a January Series A of $150 million and a May 2025 seed round, with Arena expanding from LMArena to Arena and rolling out its first evaluation product in September 2025.

Arena and Galtea—though aiming at similar AI-behavior insights—target different products: Arena provides model rankings for makers, while Galtea supplies regulatory evidence documents for compliance teams.

The analysis positions Arena within a 'Business Engineer' framework, arguing that real-session agent behavior is the key value driver for autonomy and safety.

Plans include adding more safety signals (such as refusals to harmful prompts) and expanding the index to additional models and real-world settings, updating the leaderboard signal-by-signal.

Regulatory interest is rising, with UK ICO seeking evidence on AI data protection risks from agents and a deadline for responses later this year.

Findings indicate that failures depend on the task and session length, not solely on the model, and signals draw on OpenAI and Anthropic definitions.

OpenAI models lead Arena’s preliminary alignment leaderboard, with Claude Opus 5.5 and Claude Fable also ranking among the top models.