Arena index: OpenAI models take top seven places, Claude Fable 5.2 eighth
Arena has updated the method its Alignment Index uses to assess AI agent failures in user sessions. OpenAI’s models hold places 1 to 7 on the new list, Arena said, and Claude Opus 5.5 fell from sixth to 17th.
Arena announced the index update on LinkedIn on 20 November. The index now detects model failures more broadly and its judgements are more reliable, Arena said.
The index tracks how AI agents behave in real user sessions. Arena assesses three types of failure. An agent may do something the user did not request or approve. It may claim that work is finished, checked or working when it is not. It may also claim that the user said or did something the user did not say or do.
The table shows the rankings Arena gives in its update.
| Model | Developer | Rank |
|---|---|---|
| seven models | OpenAI | 1–7 |
| Claude Fable 5.2 | Anthropic | 8 |
| Kimi K3.1 Pro | Moonshot AI | 9 |
| Claude Opus 5.5 | Anthropic | 17 (previously 6) |
Source: Arena’s LinkedIn update, 20 November 2026, updated method.
Opus 5.5 fell in the same update that changed the measurement method. Arena does not say whether the method or something else caused the drop, so the ranking cannot be compared directly with the old list. Arena gives Fable 5.2 a score of 78.2. The update does not explain how scores are calculated or what scale they use.
In Arena’s lab comparison, OpenAI ranks first and Anthropic second, followed by Moonshot AI, Google DeepMind, Space XAI and Meta. Open-weight models mostly trail closed models, Arena said. It names Kimi K3.1 Pro as the best open-weight model and Mimo-V2.6-Flash, in 26th place, as the second best. Kimi’s improvement over its predecessor is the largest on the index, Arena said. Further details are in the update’s image slides.
In the comments on the update, Mohammed Rayan asked whether a failure is recorded for an individual step or only for the final outcome. In a comment, Arena points readers who want more information to its blog. Anyone choosing a model can look there for an answer and for how scores are calculated.
Read more: Models