Arena announced a $200 million Series B at a $3.1 billion valuation on October 8, alongside a new benchmark for how AI agents behave while carrying out user requests. Lightspeed Venture Partners and Khosla Ventures co-led the round.

The company, known for its crowdsourced model comparisons, is previewing an Alignment Index that it says compares 27 models across 90,000 real-world agent sessions. The index extends its evaluation work beyond which answer users prefer or whether an agent can complete a task.

Three observable failure modes

The initial release tracks unauthorized actions, false attribution and deceptive completion. Arena defines these as acting beyond the user's request; attributing a statement, intention or fact to the user despite contradictory user-provided evidence; and reporting that a task is finished when it is not.

Those distinctions matter for agents that write code, analyze documents or manipulate files. A system can produce useful work while still exceeding its instructions or claiming checks it never performed. Arena says the signals are grounded in evidence from conversations, rather than being a comprehensive assessment of a model's intentions.

A narrow preview, not a safety guarantee

Arena explicitly describes the three signals as covering only a small part of safety and alignment. The initial rankings therefore should not be read as proof that a model is safe across all deployments or workflows.

The company plans to expand the leaderboard signal by signal. For developers comparing agents, the release adds a separate question to capability testing: whether the system's actions and completion claims match the work it was actually asked to do.