Moonshot launches Kimi K3.1 Pro, with open weights coming in two weeks: Reveneau

Why it mattersAn open-weight model that scores near Claude Opus 5.5 and GPT-6.1 Sol on coding is two weeks away. A team paying frontier prices gets a self-host option to test.

An open-weight model that targets Claude Opus 5.5 and GPT-6.1 Sol is now available on the Kimi API. Moonshot says the full weights for Kimi K3.1 Pro will be released by November 4, 2026. A team on a paid API gets a self-host option to test in two weeks. Simon Willison noted the release the same day.

Moonshot states the model is 3T-class in size. It keeps a 1 million token context window. There are two versions. Kimi K3.1 Pro is the standard model. Kimi K3.1 Pro Highspeed is 2 to 3 times faster at twice the price. The exception is input cache hits, which cost the same. Both are available in the Kimi API, Kimi.ai, Kimi Work and Kimi Code.

The pricing claim

Simon Willison reports the US dollar price from the Moonshot platform. Input is $2.00 per million tokens. Output is $10.00 per million. That matches GPT-6.1 Sol's list price. It is half the list price of Claude Opus 5.5. The Moonshot pricing page lists a cache write at $2.00 per million. A cache hit costs $0.20 per million.

Moonshot says the model used "substantially larger post-training compute" than the previous K3. The extra budget was aimed at coding, 3D design and finance.

Moonshot's own benchmark numbers

Every figure below is Moonshot's own. No independent run is available yet.

On Terminal-Bench 4.0, Moonshot reports Kimi K3.1 Pro at 65.7 percent. The company places that ahead of GPT-6 Astra at 58.2 percent, and narrowly behind Claude Opus 5.5 at 66.4 percent. On its own Kimi Code Bench 3.0 long-horizon split, Moonshot reports 61.2 percent, up from the previous K3's 55.95 percent.

For 3D design, Moonshot reports a BenchCAD score of 92.7 percent with tools and 72.8 percent without tools. It places the with-tools score close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent.

For finance, Moonshot reports 64.4 on OfficeQA Pro v2 and 58.6 on its own Frontier Finance set. Its own internal financial research score moved from 58.8 for K3 to 70.4 for K3.1 Pro. That is a lift of 11.6 points, by the lab's own measurement.

On hallucination, Moonshot says the claim-level rate dropped from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. That is the lab's own measurement on its own test set.

One independent reading comes from the Artificial Analysis leaderboard. Simon Willison reports a score of 55 there. He places it ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53. It sits behind Claude Opus 5.5 at 58. Willison calls it the first time a soon-to-be open-weight model has scored this high against the frontier set.

What a team deciding on a paid model should do

The model is on the API today, so a team can run its own evaluation set against it this week. The November 4 date for open weights is Moonshot's own promise. A team that holds that promise lightly can wait two weeks. That keeps the self-host choice open, if the API numbers hold up.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.