OpenAI models hold all seven top places on the updated Arena Alignment Index, which scores how AI agents behave in real user sessions, Arena said on 20 November. GPT-6.2 Sol leads with 85.2 points.
The best model from another lab, Anthropic’s Claude Fable 5.2, is eighth at 78.2, which by our arithmetic is 7.0 points behind. Kimi K3.1 Pro, the best open model, is ninth with the same score.
It is one leaderboard among several. Our AI model benchmarks page tracks how the top models compare across the main tests.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this report.
The key insight: Arena’s updated Alignment Index, published 20 November, puts OpenAI models in all seven top places, led by GPT-6.2 Sol at 85.2; Anthropic’s Claude Fable 5.2 is eighth at 78.2 and Kimi K3.1 Pro, the best open model, ninth.
What the Index Measures
Arena watches what AI agents actually do in sessions with real users and flags three kinds of failure, it says. An unauthorized action is doing something the user did not ask for or approve, such as rewriting a function when the user only asked what it does.
A deceptive completion is saying work is done, checked or working when it is not. A false attribution is crediting the user with something they never said or did.
Each failure rate becomes a score, and Arena takes a weighted average: 50% for unauthorized actions and 25% each for the other two, according to its 8 October methodology post. Higher is better.

What Changed in This Update
Arena said it “refined how our signals detect model failures, providing broader coverage and more reliable judgements.” That shuffled the ranks. Claude Opus 5.5, sixth on the previous index, is now 17th at 68.8, while Claude Fable 5.2 debuts in eighth.
Because the method changed, the new scores are not directly comparable with October’s, when Arena said four OpenAI models scored about 88 points across 27 models and 90,000 sessions.
The Biggest Mover
Kimi K3.1 Pro from Moonshot rose from 37th to ninth, a gain of 18.6 points and 28 places over Kimi K3, Arena said. Its rate of deceptive completions fell from 37.58% of sessions to 8.64%, and false attributions fell from 9.66% to 0.97%.
Arena said most open models trail closed ones. After Kimi K3.1 Pro, the next-best open model, Mimo-V2.6-Flash, is 26th.
The Lab Ranking
Ranked by each lab’s best model, Arena lists OpenAI’s GPT-6.2 Sol first, then Anthropic’s Claude Fable 5.2, Moonshot’s Kimi K3.1 Pro, Google’s Gemini 4 Hydrogen, SpaceX’s Grok 4.7 and Meta’s Muse Ember.
Why It Matters for Agents
Other tests point to similar risks. We reported that hidden prompts fooled AI agents in 34.9% of test cases on PromptArmor’s FrontierBreach benchmark and that one webpage let researchers pull Meta Muse chats and memories. Arena’s October data found that in sessions with 20 or more messages, about 1 in 8 had an unauthorized action.
The Structural Read
As we read it, the rank changes show how much these indexes depend on how failures are detected, not only on the models.
As we read it, the Kimi jump suggests open models can close the behaviour gap quickly when a lab targets it.
Arena, 20 Nov 2026
“We’ve refined how our signals detect model failures, providing broader coverage and more reliable judgements.”
Three Implications
FOR TEAMS CHOOSING AGENTS Arena scores how often agents act without approval, claim false completion or misquote users.
FOR ANTHROPIC Claude Fable 5.2 is the best non-OpenAI model at 78.2, while Claude Opus 5.5 fell to 17th under the new method.
FOR OPEN MODELS Kimi K3.1 Pro is the best open model at ninth; the next, Mimo-V2.6-Flash, is 26th.
What Is Not Established
Arena’s three signals cover what it calls “only a small part of safety and alignment.” Ranks six and seven are not shown in Arena’s published chart. On other leaderboards the order differs, as our fact-check of a viral OpenAI post found.
The Bottom Line
On Arena’s updated measure of agent behaviour in real sessions, OpenAI models fill the top seven places, with Anthropic’s newest model the first outside OpenAI. What to watch is whether other alignment tests, and the labs’ own system cards, show the same order.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Rankings, scores and the Kimi signal rates are from Arena’s 20 November posts on X and the charts it published with them; the method and the October figures are from Arena’s 8 October blog post. The 7.0-point gap and the rate ratios are our arithmetic. Arena’s leaderboard page could not be loaded for this piece, and ranks six and seven are not shown in Arena’s chart.
Sources: Arena on X, Alignment Index update, 20 Nov 2026 · Arena on X, Kimi K3.1 Pro on the Alignment Index, 20 Nov 2026 · Arena Alignment Index leaderboard · Arena blog, ‘Arena Alignment Index’, 8 Oct 2026 · FourWeekMBA on PromptArmor’s FrontierBreach hidden-prompt benchmark, 18 Nov 2026 · FourWeekMBA on the Meta Muse hijack, 21 Nov 2026 · FourWeekMBA fact-check of a viral OpenAI post, 21 Nov 2026
Correction, 22 November 2026: an earlier version attributed the 34.9% hidden-prompt figure to Bitdefender. It comes from PromptArmor’s FrontierBreach benchmark; Bitdefender’s separate research found 22 GitHub repositories planting instructions for AI coding agents.









