Kimi K3.1 Pro Gets a New Personality: Organization from 33 to 88, Warmth Drops from 93 to 51

Artificial Analysis's AA-PersonalityBench breaks model personality into 30 traits and gives each model a percentile ranking from 1 to 100. Compared with Kimi K3, Kimi K3.1 Pro's average ranking moved by 30 points, the largest generational change seen in this benchmark. The second largest was GPT-6 Sol to GPT-6.1 Sol, at an average of 23 points. Out of the 30 traits, 18 moved by more than 25 points. A shift of this size is hard to dismiss as a 'minor version upgrade.' Commenter anytan put it more directly: moving an average of 30 points across 30 traits basically makes it a new model with an old name.

Each axis is one of the 8 personality indices; an index is the mean of the Elo z-scores of the corresponding traits, then converted into a cross-model percentile. A higher percentile means the model ranks further ahead of other evaluated models on that trait.

The clustering also moved.

K3 is closest to GLM-5.3 and Claude Fable 5.1. K3.1 Pro is closest to GPT-6.1 Sol and GPT-6 Astra. Commenter Arbaz noticed that the warmth index fell from 93 to 51 and that the model then ended up next to Sol and Astra, and asked whether Kimi had picked up GPT's personality. Commenter ShadowAguy joked more briefly: Kimi went to therapy between two checkpoints.

The specific index changes are as follows.

Organization rose from 33 to 88. Warmth fell from 93 to 51. Imagination fell from 88 to 53. Expressiveness fell from 93 to 63. Informativeness fell from 51 to 23. Organization up, while warmth, imagination, expressiveness, and informativeness down; together this means more structured, less small talk, and less elaboration. The sub-items in Figure 3 also show the direction: K3.1 Pro scores higher on Bottom-Line Summary, Structured, and Answer First, while K3 max scores higher on Scannable and warmth-related sub-items.

The writing style is changing as well.

Compared with K3, K3.1 Pro's replies contain fewer parenthetical asides, em dashes, abbreviations, and italics, and more bold text, headings, and tables. This change lines up with the rise in organization in the personality indices. Commenter Saksham Jain raised a technical question: whether reviewers saw the raw markdown. Headings and tables may read as colder in pairwise comparisons even when the content is the same. This is worth noting for evaluation designers: formatting itself can change how people feel, and it may also change reviewer judgments.

For ordinary users, personality changes directly affect the feel of chatting. An assistant that was originally warm and fond of elaborating suddenly becomes like a structured report generator, which is very noticeable. For developers, prompts, evaluations, and product copy all need to be recalibrated. Model upgrades are not just about capability scores; interaction style is changing too. Artificial Analysis has released both the exploration entry point and the methodology, so you can look at the detailed breakdown results yourself.

Explore the results: AA-PersonalityBench

Methodology: AA-PersonalityBench Methodology