Arena raises $200M and adds real-world AI alignment measurements
Provides crowdsourced AI model performance leaderboards
The new Alignment Index examines agent behaviour during real user workflows. Its initial signals cover actions beyond the user’s permission, unsupported attribution to the user and claims that unfinished work is complete.
The launch builds on Arena’s progression from comparing preferred responses to testing factuality and multi-step agent work. It adds a behavioural view alongside capability leaderboards, helping teams examine how a model acts as well as whether it can finish a task.
Use of funds
Expand real-world evaluation of AI capabilities and alignment.