Kimi K3.1 Pro (Moonshot AI)

Summary

Moonshot AI released Kimi K3.1 Pro, a post-trained version of Kimi K3, on its API and apps. It has a 1M-token context window, and Moonshot promises full open weights by 2026-11-04. The blog post has no date. The release date of 2026-10-23 comes from Simon Willison's same-day write-up and from Pydantic AI adding the model that day. Moonshot calls it a "3T-class" model but gives no architecture details. The prices below are from Willison's post, because the blog only links to a pricing page.

Key points

• IDs: kimi-k3.1-pro, and kimi-k3.1-pro-highspeed (2–3× faster at twice the price; cached input costs the same).
• Price (reported by Willison, 2026-10-23): $2 per 1M input tokens and $10 per 1M output tokens. That is the same as GPT-6.1 Sol and half the price of Claude Opus 5.5.
• Benchmarks (Moonshot's own, reasoning effort "max"): Terminal-Bench 4.0 65.7% (Opus 5.5 66.4%, GPT-6 Astra 58.2%). BenchCAD with tools 92.7% (Opus 5.5 96.2%). Its claim-level hallucination rate with tools on Moonshot's own Kimi Hallu Bench Hard fell to 8.3%, from 23.4% for K3. Artificial Analysis gave it an index score of 55, behind only Opus 5.5 (58) (from Willison's post).
• Where to use it: the Kimi API (with zero data retention for API customers), Kimi apps, Kimi Work desktop 3.1.0 or later, and Kimi Code (/model to select it). Enterprise admins must turn it on in Kimi Business.
• Weights: not released yet. The licence is not named.

Why it matters

If Moonshot keeps its promise, this is the strongest open-weight coding and agent model so far, close to Opus 5.5 on Terminal-Bench at a lower API price. Until the weights and independent results appear, the benchmarks are only the developer's own figures.