The updated Arena Alignment Index now features OpenAI models in all seven top positions. The GPT-6.2 Sol model currently leads the ranking with a score of 85.2, followed by GPT-6 Astra, which trails by 0.6 points.

The index, maintained by Arena.ai, evaluates AI agents based on real-world interactions rather than laboratory tests. It monitors three failure types: unauthorized actions, false attribution, and deceptive task completion. The assessment utilizes data from over 195 000 agent sessions.

Anthropic’s Claude Fable 5.2 debuted at number eight, while Claude Opus 5.5 dropped to the seventeenth position following the methodology update. The revision introduced stricter signals for detecting failures, which now occur at rates generally below 5-10% for top models.

Moonshot AI’s Kimi K3.1 Pro holds the ninth spot with a score of 78.2, making it the highest-ranked open-weight model. This category includes models with publicly released parameters that users can run on their own hardware.

Arena.ai first launched the index in early October 2026, with the latest update arriving between November 19 and 20, 2026. The platform also introduced new focused indexes to broaden the evaluation of AI alignment with user intent.