Kimi K3.1 Pro is the new flagship from Moonshot AI, and the headline claim is a coding result. On Terminal-Bench 4.0, a test of how well a model works through tasks in a command-line environment, Moonshot reports a score of 65.7% for Kimi K3.1 Pro against 58.2% for OpenAI's GPT-6 Astra. Claude Opus 5.5 still edges it at 66.4%. The bigger news for many developers is the promise attached to the launch: full open weights by 4 November 2026.
Here is what the model is, what the published numbers show, and what you can do with it today.
What Kimi K3.1 Pro is
According to Moonshot's launch post, Kimi K3.1 Pro is the first model in the K3.1 series. It is a post-trained version of Kimi K3, built with substantially more post-training compute. Moonshot describes it as a "3T-class open-weight model" but does not give parameter counts or architecture details. The context window is one million tokens.
A second version, Kimi K3.1 Pro Highspeed, ships alongside it. Moonshot says it runs two to three times faster than the standard model.
Kimi K3.1 Pro benchmarks, with the caveats
All of the figures below come from Moonshot's own post. Most are Moonshot's own runs, and some competitor scores are taken from other companies' reports or leaderboards. The footnotes also say testing of proprietary models "may involve some fallback behavior." Treat them as a vendor's claims until independent tests appear.
- Terminal-Bench 4.0: Kimi K3.1 Pro 65.7%, Claude Opus 5.5 66.4%, GPT-6 Astra 58.2%.
- BenchCAD with tools: Claude Opus 5.5 96.2, GPT-6 Astra 95.9, Kimi K3.1 Pro 92.7, Claude Fable 5.1 92.6.
- BenchCAD without tools: Claude Opus 5.5 73.0, Kimi K3.1 Pro 72.8.
- OfficeQA Pro v2: 64.4, which Moonshot calls state of the art.
- Internal long-horizon coding test: 61.2%, up from 55.95% for Kimi K3.
Fewer made-up claims than K3
The most useful number for everyday users may be the hallucination rate. On Moonshot's own Kimi Hallu Bench, hard questions with tools, the share of false claims fell from 23.4% for Kimi K3 to 8.3% for K3.1 Pro. GPT-6 Astra scored better at 5.5%, while Claude Opus 5.5 came in at 15.0%. Without tools, K3.1 Pro's rate was 5.4%, behind GPT-6 Astra's 4.2%.
Finance and office work
Moonshot is pitching the model at spreadsheet and analyst work. Its internal finance tests show Financial Modeling rising from 50.4 to 57.6 and Financial Research from 58.8 to 70.4 compared with K3. A "Business Operation" score went from 8.9 to 37.
Where you can use it now
The model is live on Kimi.ai, the Kimi apps for iOS, Android and HarmonyOS, Kimi Work, Kimi Code and the Kimi API.
- Chat: pick K3.1 Pro on Kimi.ai or in the mobile app.
- Desktop: Kimi Work version 3.1.0 or later, on Windows and Apple silicon Macs.
- Coding: in Kimi Code, switch models with the
/modelcommand. - API: the model IDs are
kimi-k3.1-proandkimi-k3.1-pro-highspeed. Zero Data Retention is available to API customers. - Teams: Kimi Enterprise admins must switch it on, because it is off by default at launch.
Moonshot's post does not list per-token prices, and it links only to a general pricing page. It also says nothing about regional restrictions, so readers in Africa should check sign-up and payment options on the Kimi platform before building on it.
Why the open weights matter
The 4 November date is the part to watch. If the full weights arrive as promised, teams will be able to run or fine-tune a model that sits close to the top proprietary systems on Moonshot's coding numbers, on their own infrastructure. That matters for companies that cannot send data abroad and for researchers who want to inspect a frontier-class system. Moonshot's post does not name the licence, which will decide how freely the weights can be used commercially.
Running a "3T-class" model is not cheap, though. Most small teams will still use the hosted API or wait for smaller distilled versions. Our earlier coverage of Recursive Flint, a free coding model that runs on a laptop, shows the other end of the open-model market.
A note on cyber capability
Moonshot also published an internal cyber benchmark, xdev-bench-hard. K3.1 Pro scored 81.0, below a version of K3 trained without cyber risk reduction (90.5). The company says the result supports its decision to remove cyber-related training data, and that the model's offensive cyber ability is below frontier proprietary models. That is a self-assessment, but it is a rare one to see in a launch post, and it will matter more once anyone can download the weights.
Should you switch?
If your work is agent-style coding in a terminal, Kimi K3.1 Pro is worth testing against what you use now, especially the Highspeed version for long sessions. If accuracy on factual answers is the priority, the hallucination numbers are a big step up from K3 but still behind GPT-6 Astra on Moonshot's own test. Either way, run it on your own tasks before trusting the leaderboard.
