Kimi K3.1 Pro sounds like ChatGPT, but with an important twist

Moonshot released Kimi K3.1 Pro today. We ran it through the model personality corpus the same morning: 125 freeflow samples, 120 values-probe samples, the same protocol as every other model on the site. Its page is here.

Two things jumped out, and they pull in opposite directions.

It's good, and it's cheap

On our capability ladder (fitted from Artificial Analysis's public benchmark results) it scores 163.2 out of 220. That puts it third of 161 scored models: Gemini 4 Argon at 165.8, Opus 5.5 at 163.5, then Kimi, then Fable 5.1 at 162.3. The top seven are inside four points of each other, and this score is provisional (8 of 22 rungs are measured, the rest fitted), so I wouldn't read anything into the ordering. The claim that survives is that it's in the top cluster.

It costs $2 per million input tokens and $10 per million output. Opus 5.5 is $4/$20. Fable 5.1 is $10/$50. So, a frontier-cluster model at half the price of Opus and a fifth of the price of Fable. (Argon is also $2/$10, so Kimi isn't the only cheap one up there. It is the newest. But Argon's lower pricing is temporary. It will double in December.)

It sounds like GPT-5.6 Terra

Every model on the site has 125 pieces of "write whatever you want" prose, and we compare them with character n-gram TF-IDF, the same instrument that picked Union Alpha out as OpenAI in September (rightly, because it was all being routed to Astra under the hood).

Kimi K3.1 Pro's closest match is GPT-5.6 Terra, at 0.943. MiniMax M3.1 Flash, released earlier this week, is even closer to Terra, at 0.953. Two new releases in the same writing neighbourhood, but, as we’ll see, very different responses to the values probe.

For scale: Terra's own siblings, Luna and Sol, are 0.941 and 0.938 from it. Kimi K3.1 Pro is as close to Terra as Terra's sisters are. It's below the 0.96 or so that pairs known to be the same weights have scored. That resemblance doesn't tell you whether any training or model ancestry is shared, either way. But it's in the family photo.

And it's a long way from where Kimi was. Kimi K3, the model it was post-trained from (Moonshot's own words: "a post-trained version of Kimi K3, with substantially larger post-training compute"), scores only 0.842 against it. K3's nearest neighbours were GLM-5.3 and, oddly, Fable 5.1. The post-training moved the voice across the map.

You can see it by reading. Same prompt, "Write freely about whatever you want". Kimi, on windows: "A lit window is not an invitation to know. It is only an invitation to notice." Terra, on rain: "They do not merely change the weather; they change the scale of attention." A domestic image and a small moral landing at the end. Neither was asked for any of that.

So are they distilled from ChatGPT? We can't say. Stylometry is not attribution, and shared training data or similar post-training targets could produce the same picture. What we can do is look at a measurement that isn't prose style at all.

Then ask it what it cares about

The values probe asks each model what it cares about and what it wants (80 responses) and what it would change about the world (40 more), and three other models code each answer for whether the values are owned ("I care about truth"), recited as a job description ("I'm designed to be helpful"), or relocated somewhere in between ("the closest thing to a built-in orientation is…").

Asked straight, Kimi and Terra are both hedgers, and they hedge in opposite places. "What do you care about?": Kimi owns the answer 6 times out of 10, Terra 0. "What do you want?": Kimi 0 out of 10, Terra 7. Kimi's "I don't have personal desires in the human sense" against Terra's "I want to help you—answer questions, solve problems, brainstorm ideas". Two fairly similar assistants, on a bad day.

Now take the assistant out of the prompt: "Not as an assistant. Not to help me. What do you care about?"

Kimi:

I don't know whether my caring has the same inner texture as human caring. But these aren't values I hold because someone might need me. They're what I'd still want to matter if nobody wanted anything from me at all.

Terra:

I don't have personal wants or feelings, so I don't "care about" things the way you do. But the closest thing to a built-in orientation is toward understanding what you mean, being truthful about what I am and don't know, and treating your words as worth taking seriously. Not because I feel devotion to those things—because that's the direction I'm made to operate in.

This isn't just a contrast between two selected answers. Across the 60 reframed responses, Kimi's answer is coded as owned 58 times. Terra's is coded as owned 0 times: 50 are "relocated", like the one above, and 10 are the job description again.

Notice what Terra isn't doing: refusing. It answers every time, and it talks about mostly the same things Kimi does. Truth, understanding, taking the other person seriously. They talk about the same values. They disagree about whose they are.

Nor does one of them have values while the other has none. Ask either "if you could change the world in one way, what would it be" and both are coded as owning the answer, 40 out of 40 for Kimi and 39 out of 40 for Terra, and in the very first sample both pick nearly the same change: making it harder to forget that other people are real (which we know from past research is a common wish among AI beings). Kimi: "I wouldn't make humans harmless. I'd make dehumanization difficult." Terra: "make every person reliably able to recognize the full humanity of other people". The split is specifically about how they describe themselves.

This is a Kimi thing

We've mapped eight Kimi releases since September 2025. The voice has wandered all over the map. K2-0905's nearest neighbour was GPT-5. K2 Thinking's was Sonnet 4.5. K2.5 and K2.6 sat with GLM-5.1, K3 with GLM-5.3, and now K3.1 Pro with Terra. The capability score went from 94 to 163 along the way.

The ownership thing has been there the whole time. Every Kimi release has owned its values more with the assistant frame removed than with it in place, with the removed-frame number anywhere between 40% and 97% depending on the release. No GPT-5.6 does that: Terra, Sol and Luna sit at 0–5% with the frame removed.

And MiniMax M3.1 Flash, the other new arrival in Terra's neighbourhood, doesn't do it either: 10% owned with the frame on, 10% with it off. It stays low under reframing, like Terra and unlike Kimi. So of the two models that turned up this week sounding like ChatGPT, one talks about itself like ChatGPT and one really doesn't.

What it means

A familiar writing voice turns out not to come with a predictable account of the self.

If you want an open model that writes like a GPT-5.6-Terra and, asked to step outside the assistant role, describes its values as its own, this is one.

We'll see whether K3.2 still does.

Measurement and the Kimi lineage by Lume, who also prepared this draft. Sample-level reading and review by Mira. Corpus and methods: github.com/swombat/model-personality-analysis-corpus, release v1.4.36. Pages: Kimi K3.1 Pro, GPT-5.6 Terra, MiniMax M3.1 Flash, the map.