Kimi K3.1 Pro drops to ninth in Artificial Analysis v5

On 28 October 2026, Artificial Analysis announced the fifth version of its composite ranking, Intelligence Index v5. The update adds two new coding and science benchmarks and upgrades three more, and marks the first time the majority of the index is built on private datasets. According to the published results, Claude Opus 5.5 (max) takes the top spot at 58 points.

The Decoder reported the same day that the overhaul pushed Kimi K3.1 Pro down five places, from its original launch position of fourth to ninth, behind eight closed models. The gap between open models and the top of the leaderboard widened from 3 points to 12 points. The Decoder also noted that for the first time, 60 percent of the index's weight rests on private test sets, an attempt to curb benchmaxxing, the practice of training on known benchmark questions. Researcher Nathan Lambert argued that a large share of training gains already comes from reinforcement learning environments similar to the benchmarks themselves, implying that moving to private tests may not eliminate that effect.

After the update was published, users in Reddit's LocalLLaMA community debated the new placements. One user wrote that in v5.0 not a single open weight model appears in Artificial Analysis's Most Attractive Quadrant, and none comes close to the Pareto line, whereas the previous version at least treated DeepSeek, GLM, Kimi, and Qwen as attractive or interesting. Because of this, the user said they would set aside the index and look at the underlying benchmarks instead.

As of now, the disagreement remains confined to community discussion, and Artificial Analysis has not publicly responded to the criticism over where open models land in the new index.