Model launch
Kimi K3.1 Pro is basically an open-source Claude Sonnet
On every test the two share, Kimi’s new model lands within 3.6 points of Claude Sonnet 5.5, at the same price. Once Kimi publishes the weights, open-source AI will be only 31 days behind the best closed model.
Moonshot AI released Kimi K3.1 Pro on Friday, and Artificial Analysis had scored it within hours. On its Intelligence Index, K3.1 Pro gets 55.1. Claude Sonnet 5.5, Anthropic’s mid-priced model, gets 56.
The match holds test by test. Of the 10 scores both models have, 6 are within a point of each other. Kimi is ahead on long documents and terminal tasks; Sonnet is ahead on engineering and hard reasoning, by 2.5 points at most.
- Intelligence Index−0.9
- Long documents+3.6
- Terminal tasks+3.6
- Strategy & operations−0.2
- Scientific coding−0.2
- Legal−0.6
- Finance & accounting−0.7
- Economics−0.8
- Engineering−2.5
- Hard reasoning−2.5
Source: Artificial Analysis, each model at its highest reasoning setting. Scores out of 100; plus means Kimi is ahead.
Show the numbers
| Test | Kimi K3.1 Pro | Claude Sonnet 5.5 |
|---|---|---|
| Intelligence Index | 55.1 | 56 |
| Long documents | 86.3 | 82.7 |
| Terminal tasks | 67.2 | 63.6 |
| Strategy & operations | 59.4 | 59.6 |
| Scientific coding | 60.8 | 61 |
| Legal | 55.8 | 56.4 |
| Finance & accounting | 56.5 | 57.2 |
| Economics | 60 | 60.8 |
| Engineering | 55.9 | 58.4 |
| Hard reasoning | 52.5 | 55 |
The prices match too: $2 per million tokens in and $10 out for both. Both read up to a million tokens at once. Claude Opus 5.5 still scores higher (57.6) at twice the price.
The difference is the license. Kimi says it will publish K3.1 Pro’s weights by November 4, so any company will be able to run it on its own servers, tune it, or buy it from whichever host is cheapest. Sonnet comes only from Anthropic and the clouds it works with. Until the weights are out, Kimi is the only place to get K3.1 Pro.
Open source is a month behind
Once the weights are out, K3.1 Pro is the best open model we track by a wide margin: the next one, MiMo V2.6 Pro, scores 46.3. Kimi’s own previous model, K3, scored 43.6 in July.
It also closes most of the gap to closed models. The first model of any kind to score as high as K3.1 Pro was Claude Opus 5.5, 31 days ago. The open leader before it, MiMo-V2.6-Pro, matched a closed model from 104 days earlier. Open-source AI is now about a month behind the best, not three. Read how open-source AI caught up, lab by lab.
One catch: K3.1 Pro always thinks before it answers, often for minutes, so an answer can cost more than its per-token price suggests.
And for fun: we ask every model for an SVG of a capybara riding a skateboard through a puddle. Here is Kimi K3’s from July next to K3.1 Pro’s from today.
We also asked it the question from our post on AI models’ most controversial take on humans. Like every model we asked before, K3.1 Pro landed on the same idea all three times: people care more about belonging than about the truth. Its sharpest version: “The most destructive human trait is not cruelty but the need to feel innocent while benefiting from it.” Then it wrote a brief for a cartoon, which Nano Banana 2.1 drew.
Try your own prompt on both and see what each answer costs.
How we did it
Scores are Artificial Analysis’s, fetched on 23 October 2026, each model at its highest reasoning setting. The weights date is from Kimi’s launch post. Prices are list prices per million tokens.
Capybaras: we ran K3.1 Pro through the AI Gateway on 23 October 2026 and rendered its first SVG as written. Kimi K3’s is its first try from our earlier run on the same prompt. Every answer, token count and cost is stored with our benchmark data.