Arena Alignment Index ranks OpenAI models in all seven top spots after methodology update
OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
OpenAI holds all seven top spots, while Moonshot AI's Kimi K3.1 Pro leads open-weight models at ninth
OpenAI just swept the podium, and then some. The updated Arena Alignment Index places OpenAI models in all seven top positions, with GPT-6.2 Sol in first place at a score of 85.2.
That ranking comes after Arena.ai revised how the index detects failures. In other words, the test got harder to fool, and OpenAI still came out ahead.
What the new rankings show
GPT-6.2 Sol leads the index with its 85.2 score. Right behind it sits GPT-6 Astra, another OpenAI model, trailing by just 0.6 points.
The first non-OpenAI name shows up at #8. That slot belongs to Anthropic’s Claude Fable 5.2, which debuts on the index there.
Moonshot AI’s Kimi K3.1 Pro lands at #9 with a score of 78.2. That makes it the highest-ranked open-weight model on the board.
Open-weight models are those whose underlying parameters are publicly released. Anyone can download them and run them on their own hardware, rather than renting access through a company’s servers.
Not every Anthropic model had a good week. Claude Opus 5.5, previously the leading model, dropped to #17 after the update.
How the index works, and what changed
Most AI benchmarks ask whether a model can solve a problem. The Arena Alignment Index asks a different question: does an AI agent actually do what the user wanted?
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Arena.ai built the index around real-world agent interactions rather than lab-style quizzes.
The index tracks three specific ways agents go wrong. The first is unauthorized actions, where an agent takes steps the user never approved. The second is false attribution, where an agent misrepresents who did what or where information came from. The third is deceptive task completion, where an agent claims a job is done when it is not.
The assessment draws on hundreds of thousands of real agent sessions. One recent snapshot cited more than 195K sessions.
Arena.ai first released the index in early October 2026. The extensive update arrived around November 19-20, 2026, introducing refined signals for detecting failures.
Even with sharper detection, the leading models are not failing often. Failure rates among top models generally remain below 5-10% for the signals the index tracks.
The November update also coincided with a batch of new focused indexes from Arena.ai. Those additions broaden how the platform evaluates whether AI systems stay aligned with user intent.
Why the alignment scorecard matters
For Anthropic, the picture is mixed. A debut at #8 for Claude Fable 5.2 is respectable, but the slide of Claude Opus 5.5 to #17 shows how much a methodology change can reshuffle the board.
Kimi K3.1 Pro’s showing is notable for the open-weight camp. Landing at #9 with 78.2 puts Moonshot AI’s model within striking distance of closed, proprietary systems on alignment.
The gap remains real, though. Kimi K3.1 Pro’s 78.2 sits 7 points below GPT-6.2 Sol’s 85.2, a meaningful distance on an index where the top two are separated by 0.6 points.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

