Kimi K3.1 Pro codes almost like Claude: know 3 things first

A new Chinese AI model that is free to try scored 65.7% on a coding test, almost level with Claude Opus 5.5 at 66.4%, by its maker's own count. But before you type anything into it, there are three things to know.

The model is Kimi K3.1 Pro from Moonshot AI, a Chinese AI company, which released it on 23 October 2026 and says anyone will be able to download the full model by 4 November, according to Moonshot's announcement.

After reading, you will know what the model is good at, how its scores compare, how to try it, and what to keep out of it.

Kimi K3.1 Pro reads 1 million tokens at once and focuses on code and 3D

K3.1 Pro is an upgraded version of Kimi K3, released in July.

It can read 1 million tokens at once, tokens being the small chunks of text an AI reads, or roughly several thick books, and it takes both text and images. Moonshot built it with 2.8 trillion parameters, of which 104 billion work on each request, Notebookcheck reported.

Moonshot highlights four kinds of work the model is built for:

• Coding whole projects: including games in Unity, Unreal Engine and Godot
• 3D work: building 3D models in Blender and CAD software
• Finance and research: financial modelling and research
• Agent work: multi-step tasks the AI carries out on its own

A tool called 3D Make lets users preview and edit 3D models on kimi.ai and download them ready for 3D printing. A second version, K3.1 Pro Highspeed, gives the same answers two to three times faster, Moonshot says.

Independent tests put Kimi behind Claude Opus 5.5 but ahead of GPT-6 Astra

The 65.7% is Moonshot's own score.

The test is Terminal-Bench 4.0, in which AI does real coding work through a terminal, the screen where you type commands. Moonshot says K3.1 Pro scored 65.7%, ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Anthropic's Claude Opus 5.5 at 66.4%.

It does not win everywhere. On a science test it trails both rivals by a wide margin, VentureBeat reported.

Artificial Analysis, an independent tester, gave K3.1 Pro 55 on its Intelligence Index. Claude Opus 5.5 scored 58, and GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon scored 53, according to Simon Willison, a developer who follows AI models closely.

Willison said it is the first time a model about to be released for download has scored this close to the leaders.

It is also cheaper. Through the API, K3.1 Pro costs US$2 per million input tokens and US$10 per million output tokens, half the price of Claude Opus 5.5. Highspeed costs twice as much.

Try it free on the web and in the app

You can start in three steps.

1. Go to kimi.ai or download the Kimi app for iOS or Android, and sign in.
2. The free plan includes limited use of K3.1 Pro. Highspeed needs the Pro plan or higher, Notebookcheck reported.
3. Give it a task it is built for: a simple web page or game, a 3D model in 3D Make, or a summary of a long report.

3 things to know before you type anything in

It is free to try, but know these first.

• Chats go to Moonshot: the model is too large for a normal laptop, so free users' chats go to the servers of Moonshot, a Chinese company. Keep personal data, passwords and confidential work documents out, and note that Notebookcheck reported a zero data retention option only for API customers.
• Always check the answers: on a test Moonshot designed to catch its own mistakes, 8.3% of its statements were still wrong, down from 23.4% for K3.
• Be wary of hype: VentureBeat reported screenshots, posted by a user on X, of what the user said were instructions for a paid Kimi promotion that did not require labelling posts as paid. The document has not been verified, and Moonshot had not responded as of 25 October.

That makes independent tests a better guide than posts praising a new model.

On safety, Moonshot says it removed hacking-related data from training before the public release. On its own test of exploiting software flaws, the released model solved 81.0% of tasks, down from 90.5% for a version without those cuts but still above K3's 33.3%, The Next Web reported.

The full model will be downloadable by 4 November

Moonshot says it will release the full model files, known as weights, by 4 November 2026.

Once the files are out, a company can run the model on servers it controls, so its data does not have to leave the company or the country. At 2.8 trillion parameters, though, that needs data-centre hardware, not a laptop.

Put simply, using the app or the API is like renting the model over the internet, while downloading the weights is like buying the machine and setting it up in your own office. For most Thai users, the realistic way in is the app or the API.

What to watch: whether the model files arrive by 4 November, under which licence, how independent coding tests rate the model as more people use it, and whether Moonshot answers the promotion allegations.

FAQ

What is Kimi K3.1 Pro?

An AI model from Moonshot AI, a Chinese company, released on 23 October 2026. It is strong at coding, 3D modelling, finance and multi-step tasks, and reads up to 1 million tokens at once.

Is Kimi K3.1 Pro free?

Limited use is free on kimi.ai and the Kimi apps. The faster Highspeed version needs the paid Pro plan or higher, and the API costs US$2 per million input tokens and US$10 per million output tokens.

Is it better than ChatGPT or Claude?

On Moonshot's own coding test it beats GPT-6 Astra and nearly matches Claude Opus 5.5. An independent ranking puts it ahead of GPT-6 Astra and behind Claude Opus 5.5.

Is it safe to use for work?

Treat it like any cloud chatbot: chats go to Moonshot's servers, so keep personal and confidential data out. Companies that need their data in-house can wait for the open weights and run the model themselves.

Sources

• Moonshot AI: Introducing Kimi K3.1 Pro
• VentureBeat: Moonshot's Kimi K3.1 Pro flirts with open weights leadership amid paid influencer scandal allegations
• Notebookcheck: Kimi K3.1 Pro, free to try, hacking skills deliberately weakened
• Simon Willison: Introducing Kimi K3.1 Pro
• The Next Web: Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills