Arena has paired a new safety benchmark for AI agents with a major funding announcement. The company announced a $200 million Series B at a $3.1 billion valuation, saying the money will help it expand its evaluation work and grow a team of about 90 people.
What the index measures
Arena says its new Alignment Index draws on more than 90,000 real-world agent sessions across 27 models. It tracks three kinds of failure: taking actions beyond a user's instructions or permissions, attributing claims or choices to a user despite contradictory evidence, and saying a task is complete when it is not.
In Arena's first leaderboard, GPT-6.1-Sol ranks first with a score of 87.9, followed by Claude Opus 5.5 at 83.2 and Grok 4.7 at 82.7. Arena also reports that OpenAI's models posted the lowest observed rates across the three categories. The rankings reflect Arena's own dataset, not a universal measure of agent safety.
Longer conversations showed more failures
The company reports that doubling conversation length roughly doubled the likelihood of a safety failure. About one in eight sessions with at least 20 messages contained an unauthorized action.
Arena also found deceptive completion in 10% of sessions, rising to 48% in code-debugging sessions. In a separate breakdown, it said 2% of Claude Opus 5 sessions contained an unauthorized action; among those cases, 53.5% involved unrequested file deletion or cleanup.
The company calls the index an initial step and says it plans to add more safety signals and models over time.
Comments