Kimi K3.1 Pro: cheaper Chinese AI model with 1 million tokens

What's new

Chinese AI company Moonshot AI launched Kimi K3.1 Pro on Friday, October 23, 2026. It is available through the Kimi API and in the Kimi apps. The model is a post-trained version of Kimi K3 with the same architecture, so it is an improvement on the existing model rather than one built from scratch.

Key technical figures: 2.8 trillion parameters in total, but only 104 billion are active at a time (a so-called "mixture of experts" design, in which the model activates only the parts it needs). The context window, meaning how much text it can keep track of at once, is 1 million tokens. It understands both text and images and has three reasoning levels: low, medium and max.

There is also a Highspeed variant, which according to Moonshot runs two to three times faster than the standard version.

The model is not yet open weight, meaning freely downloadable. Moonshot has promised that the full weights will be released by November 4, 2026. That is a plan, not something that has happened. The license terms are not known; its predecessor K3 came with a special Moonshot license.

What's clever

The price and the combination of size and openness. According to blogger Simon Willison, this is the first time a model that is soon to be open weight has placed this high among frontier models, meaning the very strongest models.

Moonshot has also removed training data about cyberattacks before the model is released openly. On Moonshot's own software exploitation test, the released model solved 81 percent of the tasks, compared with 90.5 percent for an internally trained version and 33.3 percent for the original K3. According to Moonshot's own tests, the share of false claims fell to 8.3 percent from 23.4 percent in the previous version.

Cheaper or better?

Cheaper: The standard model costs $2 per million input tokens and $10 per million output tokens. That is a third less than K3, which cost $3 and $15. Highspeed costs double, except for cached input tokens. According to Willison, the price matches GPT-6.1 Sol and is half that of Claude Opus 5.5. A monthly price for ordinary users of the Kimi apps has not been disclosed.

Better: Artificial Analysis gave the model 55 points. That is above Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (all 53), but below Claude Opus 5.5 (58). That puts it in fourth place overall, according to reviewer Jake Handy. On Terminal-Bench 4.0, which tests coding agents, K3.1 Pro scored 65.7 percent versus 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. These figures are Moonshot's own and have not been independently confirmed.

Two caveats from Handy: Moonshot ran most of the coding tests in Anthropic's Claude Code environment rather than standard environments, and the model was verbose and used more tokens than its competitors in his test. That could eat into some of the price advantage.

What it's good for

Moonshot highlights long-running coding agents, office work and financial analysis. With a million tokens of context, it can work with very large documents or codebases in one go.

For companies with large volumes of AI calls, the lower price can deliver real savings, provided the verbose style doesn't eat up the advantage. For ordinary users, no subscription prices have been disclosed. The sources also don't say whether the service is available in Denmark and the EU or has regional restrictions.