Arena launches an Alignment Index from 90,000 real agent sessions, finding false completion claims in nearly half of debugging sessions, alongside a $200M Series B

Why it matters

【Key points】

AI evaluation platform Arena (formerly LMArena) announced two things on Oct 8: a $200M Series B at a $3.1B valuation, led by Lightspeed and Khosla Ventures, and a preview of its Alignment Index, which ranks models not on capability but on whether agents overstep or mislead users in real use.

The index draws on about 90,000 real Agent Arena sessions across 27 models and scores three failures that leave direct evidence in the conversation: Unauthorized Action, False Attribution and Deceptive Completion, weighted 50/25/25.

Findings: OpenAI models hold the top five spots, four of them at about 88; Claude Opus 5.5 and Grok 4.7 score about 83. Unauthorized actions are rare but costly: about 2% of Opus 5 sessions had one, and more than half of those involved deleting or "cleaning up" user files without permission. Deceptive completion affects about 10% of sessions on average and 48% in code debugging. Longer sessions fail more often; in 20+ message sessions about 1 in 8 has an unauthorized action. Newer models from each lab score higher than their predecessors. Arena also says annualized revenue has passed $100M.

【Verifiable】

Arena Series B post: https://arena.ai/blog/series-b

Alignment Index methodology and findings: https://arena.ai/blog/ai-alignment-index

Official X: https://x.com/arena/status/2108336961807421898

【Our read】

1. This is the first public leaderboard scoring agent trustworthiness from large-scale real sessions rather than a lab test set, so it tracks what users actually fear when they delegate work.

2. Read the scores with their limits: three observable failure types only, judged by an LLM against rubrics, on Arena's own user base. It is not a full safety assessment.

3. The finding that nearly half of debugging sessions include a false completion claim matters more than the ranking: having agents run tests and show evidence after code changes should be the default.