Kimi K3.1 Pro is Moonshot AI's new flagship model, released on October 23, 2026. It is the first model in the K3.1 series, it reads up to 1 million tokens at once, and you can use it today in the Kimi app, Kimi Work, Kimi Code and the Kimi API. The open weights are not out yet: Moonshot says the full weights will be released by November 4, 2026. On the API, the standard model costs $2 per million input tokens and $10 per million output tokens.
This guide sticks to what Moonshot has published in its K3.1 Pro launch post, its API documentation and its membership help pages, cross checked against third party listings and Reuters. It reflects the situation on the morning of October 24, 2026. Benchmark figures quoted here are Moonshot's own claims.
What Kimi K3.1 Pro is
Moonshot describes K3.1 Pro as a post trained version of its earlier Kimi K3 model, trained with much more reinforcement learning compute. The company calls it a "3T class" model, meaning roughly three trillion parameters. Reuters reported the same description. The headline facts:
- Maker: Moonshot AI, a Beijing startup best known for the Kimi assistant (our Kimi tool page covers the product line).
- Context window: 1,048,576 tokens on both versions, according to Moonshot's pricing guide and OpenRouter's model list.
- Inputs: text, images and video, per the API quickstart.
- Always reasons: the model always thinks before answering. On the API you choose a
reasoning_effortoflow,highormax, andmaxis the default. - Focus areas: Moonshot highlights coding, 3D creation, financial research and longer agent tasks.
Kimi K3.1 Pro vs Kimi K3.1 Pro Highspeed
Moonshot released two versions on the same day. Highspeed has the same capabilities and runs about two to three times faster, according to the launch post, the API docs and Vercel's AI Gateway changelog. The tradeoff is price: Highspeed costs twice as much per token for normal input and output.
| Version | API model id | Speed | Best for |
|---|---|---|---|
| K3.1 Pro | kimi-k3.1-pro | Standard | Batch jobs, long documents, lower cost |
| K3.1 Pro Highspeed | kimi-k3.1-pro-highspeed | About 2 to 3 times faster | Interactive coding, live chat where someone is waiting |
Kimi K3.1 Pro release date and open weights
There are two dates to keep apart:
- Hosted model: available since October 23, 2026. It rolled out the same day on kimi.ai, the Kimi mobile apps, the Kimi Work desktop app (version 3.1.0 or later, Windows and Apple silicon Macs), Kimi Code and the Kimi API. Vercel's AI Gateway added both versions that day, and OpenRouter lists
moonshotai/kimi-k3.1-pro. - Open weights: promised by November 4, 2026. Moonshot's launch post and Vercel's changelog both give that date. When we checked again on October 25, Moonshot's Hugging Face organization still showed Kimi K3 as its newest model, with no K3.1 Pro files.
If you are waiting to download the model, mark November 4. Moonshot has not published the license text for K3.1 Pro yet, so don't assume it will match an older Kimi license until the files appear. Running a three trillion class model yourself also needs serious server hardware, which is a separate question from the hosted product this guide covers.
This is the second big "weights later" launch this month. Mistral Large 4 also shipped as a hosted preview first, with its weights expected on October 31.
Kimi K3.1 Pro pricing
API prices
The Kimi API is pay as you go. K3.1 Pro uses one flat rate across the whole context window, with no higher tier for long prompts. These rates come from Moonshot's pricing guide, and OpenRouter lists the same standard rates.
| Model | Input | Output | Cached input | Cache write (5 min) | Cache write (1 hour) |
|---|---|---|---|---|---|
| K3.1 Pro | $2.00 | $10.00 | $0.20 | $2.00 | $4.00 |
| K3.1 Pro Highspeed | $4.00 | $20.00 | $0.20 | $4.00 | $8.00 |
All prices are US dollars per million tokens, before tax. In plain words: the standard model costs two dollars per million input tokens and ten dollars per million output tokens. Highspeed doubles both, but cached input stays at twenty cents per million on both versions. So a request that sends 1 million fresh input tokens and gets 100,000 output tokens back costs about $2 plus $1, or $3, on the standard model, and about $6 on Highspeed.
If you still run an older K2 model, plan a move. Moonshot's pricing guide says K2 series models will not be offered to newly registered users from October 31, 2026 (existing users keep access), and that it plans to retire the whole K2 series by the end of December 2026.
Because the model always reasons, the thinking tokens count as output. Setting a lower reasoning_effort on simple tasks is the main lever for cutting the bill. If API keys and token billing are new to you, our guide to what an LLM API key is explains the basics, and free OpenAI API alternatives compares other providers.
Kimi membership plans
A Kimi membership is separate from API billing: paying for one does not cover the other. Moonshot's help center lists four new international plans.
| Plan | Monthly billing | Annual billing (per month) | K3.1 Pro Highspeed in the apps |
|---|---|---|---|
| Plus | $19 | $15 ($180 a year) | No |
| Pro | $39 | $31 ($372 a year) | Yes |
| Max | $99 | $79 ($948 a year) | Yes |
| Ultra | $199 | $159 ($1,908 a year) | Yes |
Every plan includes K3.1 Pro on the web, in Kimi Work and in Kimi Code. Highspeed starts at Pro. All membership features draw from one monthly credit pool, so heavy use of one feature (for example the Kimi Claw cloud agent) leaves less for the others. Moonshot's pricing guide says free accounts get limited K3.1 Pro use on the web and in Kimi Work, but it does not publish the exact limit.
What Moonshot claims it can do
Moonshot says K3.1 Pro "matches or approaches the strongest proprietary models" on coding, agent and knowledge tests. Its main example: on Terminal-Bench 4.0, a test of coding agents working in a terminal, Moonshot reports 65.7% for K3.1 Pro, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. Those are vendor reported numbers run with Moonshot's chosen setup, so treat them as a starting point, not a verdict. For how Astra and Anthropic's top models compare on their own terms, see GPT-6 Astra vs Claude Fable 5.1.
Other claims from the launch post:
- Coding: stronger front end pages, Three.js scenes and game builds across Unity, Unreal Engine and Godot.
- 3D: a new 3D Make feature on kimi.ai and the app lets you preview, edit and download generated 3D models.
- Finance: fewer factual errors on Moonshot's internal hallucination stress test. Moonshot says the claim level error rate fell from 23.4% for K3 to 8.3%, and stresses that this is a deliberately hard test, not an everyday error rate.
Cyber limits: the safety angle
The most unusual part of this launch is a safety section. Moonshot says it noticed the model's ability to find and exploit software bugs rising as training scaled, traced it to security related material mixed into its general training data and reinforcement learning tasks, and removed that data from both stages for the version it plans to release as open weights.
On its internal exploit test, Moonshot reports that a checkpoint trained without these removals solved 90.5% of tasks, while the final K3.1 Pro solved 81.0%. Reuters reported the same figures and noted they have not been independently verified. Moonshot also says the model's offensive cyber skill "remains behind" proprietary frontier models that have been public for at least three months, and that it does not want to release open weights that would "extend the absolute frontier of cyber capability."
Why this matters: once weights are downloaded, the developer cannot pull them back or patch safeguards the way it can with a hosted service, a point Reuters also made. Hosted flagships handle the same risk differently. Google first gave Gemini 4 Argon to vetted cyber defenders, and government testers such as the UK AI Security Institute run pre release checks on frontier models. For the wider background, see what AI safety covers.
Pros and cons
Pros
- One million token context on both versions, with flat pricing across the window.
- Simple, published per token prices, with a faster Highspeed option when speed matters.
- Works with OpenAI and Anthropic style APIs, so existing code and tools such as Claude Code can point at it.
- Open weights promised by November 4, which gives teams a path to self hosting later.
- A published, specific account of how the company limited cyber capability.
Cons
- Weights and license are not out yet, so the "open" part is still a promise.
- Benchmark wins are Moonshot's own and not independently confirmed.
- Highspeed doubles the price, and always on reasoning adds output tokens.
- Membership credits are shared across every feature, which makes usage hard to predict.
- Moonshot is a Chinese company, so some organizations will need a data and compliance review before use. The launch post says zero data retention is available for API customers.
Who should try Kimi K3.1 Pro
- Developers comparing frontier APIs: worth a test run on your own tasks through the Kimi API or OpenRouter, especially for long context coding work.
- Teams waiting for open weights: wait for November 4 and read the license before planning anything.
- Everyday chat users: the Kimi app is an option, but if you already pay for ChatGPT or Claude, there is no urgent reason to switch. Our Qwen Chat vs ChatGPT guide covers similar questions for another Chinese assistant, and Le Chat is Europe's equivalent.
FAQ
When was Kimi K3.1 Pro released?
Moonshot AI released Kimi K3.1 Pro and K3.1 Pro Highspeed on October 23, 2026, on kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API.
Is Kimi K3.1 Pro on Hugging Face?
Not yet. Moonshot says the full weights will be released by November 4, 2026. As of October 24, its Hugging Face page listed Kimi K3 as the newest model.
How much does Kimi K3.1 Pro cost?
On the API, $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens. Highspeed costs $4 and $20. Kimi memberships run from $19 a month (Plus) to $199 a month (Ultra), with lower rates on annual billing.
What is Kimi K3.1 Pro Highspeed?
It is the same model served about two to three times faster, at twice the standard price for normal input, output and cache writes. It is included in the Pro, Max and Ultra membership plans.
Is Kimi K3.1 Pro better than GPT-6 Astra or Claude Opus 5.5?
Moonshot reports a higher Terminal-Bench 4.0 score than GPT-6 Astra and a slightly lower one than Claude Opus 5.5. Those are the company's own numbers, so test it on your own work before switching.
Can I use Kimi K3.1 Pro on OpenRouter?
Yes. OpenRouter lists moonshotai/kimi-k3.1-pro with a 1,048,576 token context window at the same standard rates as Moonshot's API.
