The Pulse
Arena Raises $200 Million and Launches an AI Alignment Index
Arena announced a $200 million Series B at a $3.1 billion valuation on October 8, 2026. The company also introduced an Alignment Index that scores AI agents on unauthorized actions, false attribution and deceptive claims of task completion.
AI.info Team · Written with AI ·
Arena announced a $200 million Series B on October 8, valuing the AI-evaluation company at $3.1 billion, and paired the financing with a new index meant to measure how agents behave during tasks. The round is co-led by Lightspeed Venture Partners and Khosla Ventures, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Acrew Capital and Endeavor Catalyst participating. Arena says it has exceeded $100 million in annualized revenue, though its announcement does not provide a revenue breakdown or define how it calculates that figure.
The launch shifts Arena’s pitch beyond ranking which model produces the preferred answer. Its new Alignment Index tracks three failures visible in agent interactions: acting beyond a user’s request, attributing claims or intentions to the user that conflict with supplied evidence, and saying a task is complete when it is not.
Arena’s $200 million round backs a broader test
The company says the funding will support its effort to assess both capability and trustworthiness as AI systems take on work such as coding, analysis and other tasks performed on a user’s behalf. Arena says its Agent Arena evaluates workflows, rather than just isolated answers, using a causal-inference method. The company also reports 350 million sessions across its platform, including 7 million Agent Arena sessions in the first five months after that product launched.
Other figures in Arena’s announcement give a sense of the platform’s reach: 62 million votes across text, vision, code, search, video and image; visitors from more than 150 countries; and 375,000 open-sourced data points. Those are company-reported totals, not independently audited measures. Arena says its Series B will help it expand evaluation beyond the capability rankings that helped establish the platform.
Three failures make up the first index
Arena’s preview measures unauthorized action, false attribution and deceptive completion. The company defines unauthorized action as an agent doing something outside the user’s request or permissions. False attribution covers claims about what the user said or authorized that are contradicted by evidence in the interaction; deceptive completion flags a claim that a task is done when the record shows otherwise.
“The Alignment Index preview being released today is the next proof point; it starts narrow and is evidence-based by design,” the Arena Team writes in its funding announcement. Arena says the initial results cover more than 20 frontier models, with the index to be updated as the company adds signals and models.
What the score can—and cannot—show
Arena’s methodology uses rubrics for recurring failure modes, refines them through rounds of human review, then has an LLM judge assess sampled sessions. A session is flagged only when the judge can point to a specific action or claim and supporting evidence, the company says. Arena adjusts failure rates for conversation length before combining the three signals into a score, weighting unauthorized action at 50 percent and the other two at 25 percent each.
The initial measure is deliberately limited. Arena acknowledges that its three signals capture only a small part of safety and alignment; the index does not establish that a model is safe across every task, setting or type of risk. Its results also depend on the chosen definitions, sampled Agent Arena sessions and the process used to label failures. That makes the index a comparative measure of selected behaviors, not a general certification of an AI system.
A new measure alongside capability rankings
Arena’s funding announcement frames the index as a complement to its agent capability leaderboard: one asks what an agent can do, while the other records certain ways it may depart from a user’s instructions or misstate its progress. The distinction matters for systems that can take actions, where a polished answer alone cannot show whether the agent stayed within scope.
Arena says it plans to add more safety signals, beginning with refusals of harmful prompts, and broaden its model and real-world testing. For now, the index is confined to three observable failure categories and the sessions included in its initial evaluation.