Kimi K3.1 Pro ranks fourth on Artificial Analysis Intelligence Index

Artificial Analysis says its Intelligence Index v4.3.2 incorporates 10 evaluations.[2] Kimi K3.1 Pro earned 55 points, compared with 44 for Kimi K3.[1] The three models ahead of it were Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.[1] Artificial Analysis also said K3.1 Pro would be the highest-scoring open-weights model once its weights were released.[1] For teams, the overall rank is one comparison point; it does not replace checking the evaluations most relevant to a particular workload.

On GDPval-AA, Kimi K3.1 Pro placed third with 1,846 Elo, behind Claude Opus 5.5 and Muse Wonder Full.[1] Its 67% Terminal-Bench 4.0 score was the highest among the models in Artificial Analysis’s results, up from 13% for Kimi K3.[1] The model scored 73% on AutomationBench-AA, compared with 58% for Kimi K3.[1] Other reported gains were 47% to 52% on Humanity’s Last Exam, 23% to 25% on CritPt, and 59% to 61% on SciCode.[1] These benchmark results offer specific comparisons, rather than evidence about every real-world task.

Moonshot says Kimi K3.1 Pro and K3.1 Pro Highspeed are available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API.[3] It says the Highspeed version delivers two to three times the speed of K3.1 Pro.[3] Artificial Analysis lists API prices of $2 per million input tokens and $10 per million output tokens, 33% below Kimi K3’s rates.[1] It puts the average cost of running an Intelligence Index task at $1.98, roughly equal to Kimi K3’s.[1] Teams comparing costs should distinguish token rates from the reported cost of a benchmark task.

Artificial Analysis lists K3.1 Pro as a 2.8-trillion-parameter model with 104 billion active parameters and a one-million-token context window.[1] It also says the model’s weights had not yet been published.[1] Moonshot says it plans to release the full weights by November 4, 2026.[3] Artificial Analysis says the gap between leading open-weights and proprietary models would narrow from 12 points to three.[1] Organizations considering self-hosting can treat the stated release date as a planning marker, while checking the model’s access and evaluation results against their needs.