Kimi K3.1 Pro lands on the API, open weights promised by November 4

On October 23, 2026, Moonshot switched Kimi K3.1 Pro on through its API. You can’t download the weights yet. The company has committed in writing to publishing them by November 4, 2026 at the latest, in these words: “Full model weights will be released by Nov 4, 2026”. Until then, this open model-to-be works like any closed one: an API key goes in and an invoice comes out. Simon Willison relayed the announcement and tested the model the same day.

Pricing pitched at the top of the market

Moonshot charges $2 per million input tokens and $10 per million output tokens. That’s exactly what GPT-6.1 Sol costs, and half the price of Claude Opus 5.5. A Highspeed variant answers 2 to 3 times faster for double the price. The exception is cached input tokens, which cost the same as on the standard version. So you pay extra for speed on what the model writes, not on what it has already read.

The model offers three reasoning levels: low, medium and max. Its API is OpenAI-compatible. Simon Willison queried it with the llm openai endpoint command of his LLM tool, which arrived in version 0.32 in August. In practice, if your code already talks to an OpenAI-style API, trying Kimi mostly means swapping a URL and a key. That’s less of a migration and more of a find-and-replace.

Scores brushing the frontier

The numbers come from two places, and you should keep them apart. Moonshot’s own benchmarks put Kimi K3.1 Pro roughly level with Opus 5.5 and GPT-6 Astra: ahead on a few tests, close behind on the rest. A vendor coming out on top of its own benchmarks is about as newsworthy as a chef announcing that their soup is excellent.

The next figure tells you more, because it doesn’t come from Moonshot. Artificial Analysis gives Kimi K3.1 Pro a score of 55. That puts it ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53, and behind only Claude Opus 5.5, at 58. According to Simon Willison, this is the first time a model with open weights, or one about to have them, has scored that high against the best models of the moment.

He also gave the model his ritual test, an SVG of a pelican riding a bicycle, at all three reasoning levels. He then ran the same exercise with Kimi K3 for comparison. It isn’t a benchmark, but it’s one of the few tests no vendor has any reason to game, at least until someone launches a market for cycling pelicans.

What November 4 changes, and what it doesn’t

For twelve days, the only thing separating Kimi K3.1 Pro from its closed rivals is a promise with a date on it. If Moonshot keeps it, the comparison won’t just be about price per million tokens. It will also be about not paying per million tokens at all, as long as you have the hardware to run the model. The announcement says nothing about what that hardware would be.

It doesn’t give the license, the model size or the terms for commercial use either. “Full weights” tells you what will be published, not what you’ll be allowed to do with it. Until those terms are public, don’t build an in-house hosting plan on the November 4 date alone. And nothing stops Moonshot from keeping the Highspeed variant as a paid service: the weights come as a download, the speed doesn’t.

Until November 4, Kimi K3.1 Pro is open in the “coming soon” sense: you reach it with an API key, just like every model it wants to compete with. Once the weights are out, it’s everyone else’s price list that will have some explaining to do.