Moonshot AI released Kimi K3.1 Pro on October 23, 2026, and it's the first model in the company's new K3.1 series. It builds on Kimi K3, an open-weight model with nearly 3 trillion parameters, and adds a lot more post-training on top. The headline claim is coding, where Moonshot says K3.1 Pro lands within a point of Claude Opus 5.5 on Terminal-Bench 4.0, and the open weights are due by November 4. If you've been following the back-and-forth between open and closed models, this is one of the bigger open releases of the fall, so here's what's in it, what it costs and how to get your hands on it.
What Kimi K3.1 Pro actually is
K3.1 Pro is a post-trained version of Kimi K3. According to Unite.AI's writeup of Moonshot's release notes, Moonshot used much larger post-training compute and large-scale reinforcement learning across hundreds of domains, and it calls the result its most capable 3T-class open-weight model so far. It runs in thinking mode only, so every request reasons before it answers, and you can set the reasoning effort to low, high or max (max is the default). It also takes image and video input, which matters if you want it to read screenshots, diagrams or short clips.
The context window is 1 million tokens, which is enough to drop in a large codebase, a long contract or a pile of meeting notes in a single prompt. Moonshot also shipped a sibling called K3.1 Pro Highspeed, which it says runs 2 to 3 times faster with the same capabilities. If you're new to the difference between models you can download and models you can only rent through an API, our explainer on open-weight vs closed AI models covers the tradeoffs in plain terms.
The benchmark numbers, with the right caveats
Every number below comes from Moonshot itself, so treat it as the company's own scorecard until independent testers weigh in. On Terminal-Bench 4.0, which tests AI coding agents working in a terminal, Moonshot reports 65.7% for K3.1 Pro, 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, as The Next Web reported. On BenchCAD with tools, K3.1 Pro scored 92.7% against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra, so the closed leaders still hold an edge on some tasks.
The most interesting number for everyday users is about made-up facts. On Moonshot's internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% on K3 to 8.3% on K3.1 Pro. Moonshot points out that this test is built to provoke errors on purpose, so those rates don't describe normal use, but the direction of travel is a good sign for anyone who has been burned by a confident wrong answer. Moonshot also reports gains on its own finance benchmarks, including 7.2 points on financial modeling and 11.6 points on financial research compared with K3.
Pricing and plans
Through the API, OpenRouter lists K3.1 Pro at $2 per million input tokens and $10 per million output tokens, with a 1 million token context. The Highspeed endpoint costs $4 per million input tokens and $20 per million output tokens. That pricing puts it well below most closed frontier models for heavy coding or document work, which is a big part of why developers pay attention to Kimi releases.
In Moonshot's own apps, K3.1 Pro requires the Plus plan or higher, and the full 1 million token context is reserved for Pro and above. The Highspeed version needs Pro or above. Moonshot also offers zero data retention for API customers, which is worth checking if you plan to send it anything from work.
How to try Kimi K3.1 Pro today
You have 4 easy routes. The simplest is the web app at kimi.ai or the Kimi mobile apps on iOS, Android and HarmonyOS. Developers can switch to it inside Kimi Code with the /model command, and the Kimi Work desktop app supports it from version 3.1.0 on Windows and Apple silicon Macs. For the API, the model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
One practical tip from Moonshot's docs is easy to miss. A mistyped model ID falls back to K2.8 Preview without thinking, so if your results look oddly shallow, check the ID before you blame the model. The docs also suggest starting a new session instead of switching models halfway through a conversation. If you want a starting point for testing, our list of ChatGPT prompts that make AI more useful works just as well on Kimi, and comparing answers side by side is the fastest way to learn where each model shines.
The open weights and the safety tradeoff
Moonshot says the full K3.1 Pro weights will be published by November 4, 2026. Before that release, the company says it removed cyberattack-related material from training. On its in-house test of exploiting real software flaws, a version trained without those cuts solved 90.5% of problems, while the released K3.1 Pro solved 81%. The Next Web notes the obvious limit here, which is that open models can't be recalled once people download them, and safety researchers have warned that safeguards on open-weight models tend to lag behind closed ones.
Running the full model yourself will be a serious hardware project. A model of nearly 3 trillion parameters needs data center class GPUs, so most people will stick with the app or the API. If you liked our look at running Muse Spark Open on a Mac, smaller community-made versions may show up after November 4, so keep an eye on what hardware they actually need before you clear space on your drive.
Where K3.1 Pro fits
For most people choosing an assistant, the big 3 are still the default, and our ChatGPT vs Claude vs Gemini comparison is the place to start. K3.1 Pro is most interesting for developers and teams who want near-frontier coding at lower cost, a huge context window for long documents, and the option to self-host later. It also fits the bigger shift we covered in the biggest AI trends of 2026, where open models keep closing the gap a few months after each closed release.
If you're testing K3.1 Pro, the best way to figure out whether it's worth switching is to compare notes with people running the same prompts. The AI Trends Tribe is full of builders and tinkerers who post their side-by-side tests, prompt tricks and benchmark skepticism, and it's an easy place to ask whether the open weights are worth waiting for. And if you'd rather talk models in person, follow the Tribe Events calendar on Luma for AI deep dives, demo nights and builder dinners around the Bay Area.
