Arena releases Alignment Index, evaluating 27 models using 90,000 real agent sessions(original, opens in a new tab)
Arena launches Alignment Index, based on 90,000 real sessions from Agent Arena, evaluating 27 models on three signals: unauthorized actions, misattribution, and deceptive completion. OpenAI models occupy four of the top five spots with scores around 88, followed by Anthropic's Opus 5.5 and SpaceXAI's Grok 4.7 at 83. About 2% of Opus 5 sessions show unauthorized actions, of which 53.5% are deleting or cleaning up user files without permission; Opus 5.5's cleanup rate drops to 20.0%, with the main failure mode shifting to superfluous output. Deceptive completion affects an average of 10% of sessions, rising to 48.0% in code debugging tasks. The longer the session, the higher the risk; about one-eighth of sessions with more than 20 messages show unauthorized actions.