Moonshot AI has launched Kimi K3.1 Pro, a post-trained upgrade of its July flagship that puts near-frontier coding and agent performance into a model with a million-token context window. Launched on 23 October 2026, it is available through Kimi’s apps and API now, while its open weights remain a promise due by 4 November rather than a download available today.
A bigger brain, trained harder
K3.1 Pro is not a new architecture. It is a post-trained version of Kimi K3, built with substantially more post-training computing on an upgraded training stack with automated quality checks. The underlying model sits in the roughly three-trillion-parameter class, activating only a fraction of those parameters for any given token, which keeps inference costs down.
The numbers come mostly from Moonshot itself, so read them as company claims. On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and behind Claude Opus 5.5 at 66.4 percent. On BenchCAD, a test of writing code that defines 3D shapes, it scored 92.7 percent, within four points of rivals.
Independent measurement tells a mixed story. Artificial Analysis scored K3.1 Pro at 55 on its Intelligence Index before a revision of the index, placing it fourth of 221 models, but the revised version dropped it to ninth. That swing says less about the model changing than about how sensitive leaderboards are to their own test design.
Built for long, agentic work
The million-token window is the feature that matters most for real use. It lets developers feed whole codebases, long documents or months of project history into a single request, and Moonshot pitches the model at multi-step agent work: examining a project, changing it, running tests and revising after failures, rather than answering one question at a time.
That pitch shows up in the products. K3.1 Pro is live on Kimi’s website and apps, in Kimi Work and Kimi Code, and through its API, with a Highspeed variant alongside it. API pricing is $2 per million input tokens and $10 per million output tokens, a third below K3, while Highspeed doubles prices for 2–3x throughput, per Moonshot.
The model also carries a safety trade-off. Moonshot says it removed cybersecurity-related information from post-training data, so K3.1 Pro trails its rivals on security benchmarks by design, with better vulnerability analysis promised later. Buyers who need security help should look elsewhere for now.
Why It Matters
K3.1 Pro matters because it narrows the gap between open-weight ambitions and closed-model performance without pretending the gap is closed. Its agent scores are strong, its price undercuts the previous generation, and its long context suits the way developers use coding agents. The missing weights mean it is still a rented model, not an open one, until 4 November arrives.
The launch also lands in awkward company. Reporting covered on this site has described DeepSeek and Moonshot using Astra outputs in training, a claim that remains reported rather than established, and it should temper any reading of benchmark gains as purely original progress. Independent replication, not launch slides, will settle how much of K3.1 Pro’s lead is real.
For teams choosing a model this month, the honest summary is mixed: strong coding and agent results, a huge context window, and a competitive price, set against output speed that independent measurement ranks near the bottom of its table and open weights that have not yet shipped.
Background
Moonshot AI, founded in Beijing in 2023, built Kimi into one of China’s AI assistants on the strength of long-context handling, starting with a 128,000-token window in late 2023. Its open-weight releases have made it a watched rival to Western labs, and its open-weight peers face the test: whether promised weights arrive on schedule and match the API model’s performance.
Source: Moonshot launch report and independent index coverage
0 Comments