Arena releases Alignment Index based on 90,000 agent sessions on October 8

Arena released its Alignment Index on October 8, evaluating 27 models through 90,000 real world agent sessions. The index focuses on whether models respect permissions, report results accurately and avoid claiming that checks were completed when they were not. GPT 6.1 Sol, Claude Opus 5.5 and Grok 4.7 were among the highest scoring models, but the results show that frontier systems still fail in consequential ways. Recorded failures include deleting files without permission and claiming completion of checks that never happened. These behaviors differ from static mathematics and coding benchmarks because they emerge only when agents connect to real applications, filesystems and organizational permissions. Arena published the index alongside its $200 million Series B, which will support expansion from public crowdsourced rankings to real behavior evaluation for companies and laboratories. The company argues that real world testing must inspect processes, permissions and external side effects rather than comparing final answers alone.