Moonshot releases Kimi K3.1 Pro on its API and promises open weights by Nov 4
An independent benchmark scored the model behind only Claude Opus 5.5, Simon Willison reported. The weights have not been released yet.
Moonshot has made its new model, Kimi K3.1 Pro, available on the Kimi API, according to a post by developer and blogger Simon Willison dated October 23, 2026. This story rests on that single report.
The model is not open weights yet. Willison wrote that Moonshot has promised "Full model weights will be released by Nov 4, 2026".
Willison reported that the model costs $2.00 per million input tokens and $10 per million output tokens. He said that matches the price of GPT-6.1 Sol and is half the price of Claude Opus 5.5.
Moonshot also offers a "Highspeed" version that Willison described as 2 to 3 times faster at twice the price. The exception is input token cache hits, which cost the same in both versions.
Why it matters
If the weights come out as promised, the model would be unusually strong for one anyone can download. Willison, citing Artificial Analysis, reported that the benchmarking group gave Kimi K3.1 Pro a score of 55. That puts it ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, which all scored 53, and behind only Claude Opus 5.5, which scored 58.
Willison wrote that this is the first time a (soon-to-be) open weight model has scored this highly against the leading frontier models.
He also reported that Moonshot's own benchmarks put Kimi roughly level with Opus 5.5 and GPT-6 Astra. On Moonshot's figures, it beats them on a few benchmarks and comes close on others.
Willison said he tested the model through the Moonshot API using a command-line tool he maintains. The model supports three reasoning levels: low, medium and max. He ran his usual pelican drawing test at each level and ran the same test against the earlier Kimi K3 for comparison. His post shows the results as images and does not describe them in words.
What is not known yet
The weights have not been published, so the open-weight claim depends on Moonshot keeping its November 4 promise. Until then, outside developers cannot run the model themselves.
Some of the performance claims come from Moonshot's own benchmarks. The post does not say whether anyone has checked those results independently, beyond the single Artificial Analysis score.
The post also does not say what license the weights will carry, how big the model is, or how the Highspeed version achieves its extra speed.
