LMArena Alignment Index update: OpenAI sweeps top seven, Kimi K3.1 Pro jumps to 9th
LMArena updated its Alignment Index on November 21. The index measures how well agents behave in real user sessions, detecting model failures across three dimensions: unauthorized actions, false completions, and misattribution. OpenAI dominated this round, sweeping all seven top spots, with GPT-6.2 Sol leading at 85.2; Anthropic's Opus 5.5 fell from 6th to 17th under the new methodology, while Claude Fable 5.2 entered the ranking for the first time at 8th.
Confirmed
- GPT-6.2 Sol ranks 1st on the Alignment Index with a score of 85.2, and OpenAI takes all top seven spots
- Kimi K3.1 Pro scored 78.2, jumping from 37th place in the Kimi K3 era (59.6 points) to 9th—a gain of 18.6 points and 28 positions, the biggest improvement this round
- All three failure signals for Kimi K3.1 Pro dropped significantly, with the false completion rate down to 8.64%
- On the lab leaderboard, Moonshot ties with Anthropic for 2nd place behind OpenAI thanks to Kimi K3.1 Pro's 78.2, with confidence intervals overlapping those of Claude Fable 5.2
- Claude Fable 5.2 debuts at 8th; Opus 5.5 drops from 6th to 17th under the new evaluation methodology
Why it matters
- The Alignment Index evaluates agent alignment behavior based on real user sessions, making it closer to real-world deployment risks than static benchmarks; ranking shifts offer useful reference for assessing model trustworthiness
- Kimi K3.1 Pro's major progress shows Moonshot's iterative gains on safety behaviors such as unauthorized actions and false completions, making it one of the best-aligned models from a Chinese lab
- OpenAI's clean sweep and the ranking divergence between Anthropic's newer and older models reflect significant differences in alignment performance across model versions
2026-11-21 ~ 2026-11-22 · 6 related posts
Primary sources
- [source] Arena Alignment Index update: OpenAI sweeps top seven, Claude Fable 5.2 debuts at #8 — arena · 2026-11-21
- [source] Moonshot ties Anthropic for #2 lab on Arena Alignment Index, tops Chinese labs — arena · 2026-11-21
- [source] Kimi K3.1 Pro jumps from #37 to #9 on Arena Alignment Index, deceptive completion down to 8.6% — arena · 2026-11-21
- Kimi K3.1 Pro posts biggest alignment gain on Arena Index, deceptive completion drops to 8.6% — arena · 2026-11-21
- Arena's Alignment Index: OpenAI sweeps all seven top spots, Opus 5.5 falls to #17 — BorisMPower · 2026-11-22
- OpenAI Sweeps Top 7 Spots on Updated Arena Alignment Index; Kimi K3.1 Pro Best Open Model — gdb · 2026-11-22