Kimi K3.1 Pro: What Changed, Prices and Business Use (2026)

Quick answer: Kimi K3.1 Pro is the new version of Moonshot AI's Kimi K3 model, released on 23 October 2026. It keeps the same 2.8 trillion parameter architecture and 1 million token context window, but it was retrained for coding, agent tasks and long documents. It costs USD 2 per million input tokens and USD 10 per million output tokens, a third less than Kimi K3, and Moonshot says the full weights will be published by 4 November 2026. On coding benchmarks it now sits next to Claude Opus 5.5 and ahead of GPT-6 Astra, according to Moonshot's own numbers.

We wrote about Kimi K3 in August when its weights went open source. Since then it became the model many European companies ask us about first, because it is cheap and can run on your own servers. This guide covers what K3.1 Pro changes, what it costs for real tasks, how to access it today, what "open weights by 4 November" means in practice, and where it fits next to Claude, GPT and Gemini for a business in 2026. We build AI agents for companies across Europe and test every major model on client data, so the opinions below come from use, not from the launch post.

Kimi K3.1 Pro in five facts

• Released 23 October 2026 by Moonshot AI (Beijing) on the Kimi app, Kimi Work, Kimi Code and the Kimi API. Model IDs: kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
• Same model, new training. 2.8 trillion total parameters, about 104 billion active per token, mixture of experts, 1 million token context, text and image input. The architecture is Kimi K3's; the difference is a much larger post-training run across hundreds of task types.
• Cheaper than its predecessor. USD 2.00 input and USD 10.00 output per million tokens, cached input USD 0.20. Kimi K3 was USD 3.00 and USD 15.00. The price is flat across the whole 1 million token context, with no long-context surcharge.
• A faster variant. K3.1 Pro Highspeed runs two to three times faster, per Moonshot, at double the price: USD 4.00 input, USD 20.00 output.
• Open weights by 4 November 2026. Moonshot promised the full weights within two weeks of launch. As of 27 October they are not published yet, and the license has not been announced; K3 shipped under the Kimi K3 License.

What changed compared to Kimi K3

Three things, all from the retraining rather than from new architecture.

1. Agent and coding work. On Terminal-Bench 4.0, a test where the model has to finish real tasks in a terminal, K3.1 Pro scores 65.7 percent against 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5. On SWE-Marathon, a long software task benchmark, it leads with 51.7 percent. On Moonshot's own long-horizon coding test it went from 55.95 to 61.2 percent.
2. Fewer invented facts. Moonshot's hallucination test at claim level dropped from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. For business use this matters more than any coding score: an agent that reads your contracts and answers customers has to stop making things up.
3. Office and finance tasks. Financial modelling, research and "business operation" tasks improved by 7 to 12 points, and the model reads spreadsheets and documents more reliably (OfficeQA Pro 64.4 against 62.2 for Claude Opus 5.5).

Where it still trails: hard science and maths reasoning (HLE-Diamond 39.4 percent against 60.9 for Claude Opus 5.5 and 65.8 for GPT-6 Astra), ProgramBench (31.5 against 65.0 for Claude) and image understanding (MMMU-Pro 84.1 against 87.7). If your use case is deep research or multimodal analysis, the US models are still ahead. If it is agents, code and documents, the gap has closed.

All numbers above are Moonshot's own, run with the Claude Code harness at maximum effort, with competitor scores taken from their public reports. Treat them as a direction, not a verdict, until independent leaderboards catch up.

What a task actually costs

Prices per million tokens are abstract, so here is the same thing in tasks. One token is roughly three quarters of an English word.

For comparison, Google's Gemini 4 Argon, which reached general availability on 13 October, is listed at the same USD 2 and USD 10 until the end of November, with reports that the price doubles on 1 December. Claude Opus 5.5 and GPT-6 Astra sit several times higher per token. That is why model usage is a small line in an agent project: in our AI agent cost guide we budget EUR 20 to 150 a month for a typical agent, and K3.1 Pro lands at the low end of that range.

How to use it today

• Chat and documents: kimi.com, the iOS and Android apps, and the Kimi Work desktop app (version 3.1.0 or later).
• Coding: Kimi Code, switch with the /model command.
• API: the Kimi API is OpenAI-compatible, so any tool built for GPT models works after changing the base URL and model name. Access opens after a minimum USD 1 top-up, and a zero data retention option exists. Pricing and docs are on platform.moonshot.ai.
• Through a router: OpenRouter lists it as moonshotai/kimi-k3.1-pro, useful if you want one key for several models.
• Enterprise: in Kimi Enterprise the model is off by default; an administrator has to enable it.

Open weights: what "by 4 November" means for you

Open weights mean you can download the model and run it on hardware you control, which solves the data question for many European companies: nothing leaves your servers. Two cautions before you plan on it.

• The hardware is serious. If K3.1 Pro ships in the same format as K3, the weights are about 1.56 TB on disk and need a multi-node GPU cluster with 2 TB or more of combined GPU memory. That is a six-figure setup or a rented cluster, not a single server. For most companies the sensible path is the API with zero data retention, or a smaller open model on a rented GPU.
• The license is not published yet. K3 used Moonshot's own license. Check the model card on release before building a product on it. Moonshot also said cyber-related training data was removed from the open-weight version.

An app that calls the Kimi API, on the other hand, needs ordinary hosting. We run that kind of deployment every week: for example our hosting brand neurhost hosts Next.js and Node applications that talk to several model APIs, with daily backups and a named engineer, for companies that do not want to manage servers.

Where K3.1 Pro fits for a business

Our rule after a year of agent projects: pick the model per job, on your own data, and keep the option to switch. In that light, K3.1 Pro is a strong candidate for these jobs.

• Sales outreach and research at volume. Reading a company website and CRM notes, then drafting a personalised first email, is exactly the kind of long-context, high-volume task where a cheap, accurate model wins. Our B2B sales unit IT SalesaaS runs this kind of pipeline for IT and software companies, and the model behind it is chosen by cost per qualified lead, not by brand.
• Document-heavy back office. Invoices, contracts, tenders, long email threads: the 1 million token window with flat pricing means you can hand the model the whole file instead of chunking it.
• Coding agents and internal tools. With Terminal-Bench and SWE-Marathon scores next to Claude, it is good enough for internal automation scripts and integration work at a fraction of the price.
• Conversational products with many users. A product like our AI twin platform TalkMie, where a digital twin answers hundreds of chats a day in a person's voice, lives or dies on per-message cost and on not inventing answers. The hallucination improvement is the number we watch there.

Where we would still choose Claude or GPT: hard analytical reasoning, anything multimodal beyond simple image reading, and cases where a US vendor is a procurement requirement. Where we would choose Gemini: workloads already inside Google Workspace. Read our guide to AI agents for how we decide, and our AI integration consultancy page for how the model comparison is run on your data.

Should you switch?

If you already use Kimi K3 through the API: yes, today. Same architecture, same context, lower price, fewer hallucinations; change the model ID and rerun your tests. If you use another vendor and the job is agents, code or documents: run a two-week comparison on 50 real tasks before deciding. If the job is research or multimodal analysis: wait for independent benchmarks and the open-weight release. We run these comparisons as part of a fixed-price AI agent project, with a free one-week prototype on your data; see prices or write to info@buinsoft.com.

Last updated 27 October 2026. Prices and benchmark figures are from Moonshot AI's launch materials and API pricing page as reported on 23 to 26 October 2026; they may change. Buinsoft prices exclude VAT.