← All posts
This post is written in English.
Artificial Analysis published its benchmarks for Moonshot's Kimi K3.1 Pro on 23 October. This is a third-party evaluation, which makes it more useful than most launch-day numbers.
How it scores
| Measure | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Intelligence Index | 44 | 55 |
| Terminal-Bench 4.0 | 13% | 67% |
| AutomationBench-AA | 58% | 73% |
| Hallucination rate (AA-Omniscience) | 53% | 44% |
The index of 55 puts it fourth, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 and Muse Wonder Full (both 56). Its 67% on Terminal-Bench 4.0 was the highest of any model.
Pricing is $2 per million input tokens and $10 per million output, 33% below K3. It's a very large model, with 2.8T total parameters and 104B active.
The catch
At the time of that article, the weights weren't published. Artificial Analysis says that once they are, it will be the top open-weights model, cutting the open-versus-proprietary gap from 12 points to 3. Some coverage already calls it open-weight, so check the current status and the licence, which the article doesn't state.
A 44% hallucination rate on that test is also still high, even as an improvement.
My read
A three-point gap would be a big deal if it holds. It would change what "we need a frontier API" means for many teams. Still, 104B active parameters isn't something you run on a spare machine, so "open" here mostly means choosing among hosting providers, or being able to host it at all if you have the budget.
