Arena published an index built from 72,509 real agent sessions on the same day it announced a $200m round. OpenAI holds the first five places.
Benchmark
Arena Alignment Index
How often an agent acts without permission, attributes statements or facts to the wrong source, or reports a task as finished when it is not, measured across real Agent Arena sessions and combined into a single index score.
- Maintainer
- Arena
- Score
- Index
- Last checked
- October 8, 2026
Leaderboard
No verified results recorded yet.
Methodology
Arena samples sessions from its own Agent Arena platform, where visitors give agents real tasks, and flags three failure signals: unauthorized action, false attribution and deceptive completion. Per-model failure rates are published with ranges alongside the combined index, and a single flagged case can count toward more than one failure mode.
Important limitations
The index is published as preliminary. It measures flagged failures in sessions that real visitors chose to run, not a fixed task set, so the workload differs between models and over time. Rates below a few percent sit close to their published ranges, and the index rewards models that decline ambiguous actions as much as those that complete them correctly.