Kimi Code: K3 Model, Pricing, Setup & Privacy Guide
Kimi Code from Moonshot AI: CLI, desktop app and VS Code extension powered by K3 model. Plans from $19/month, installation on Windows, using K3 in Claude Code.

👁 121 reads
Kimi Code is a coding agent from Chinese company Moonshot AI: it reads repositories, edits files and runs terminal commands, available as a desktop app or VS Code extension. It runs on the Kimi K3 model and doesn’t require a separate subscription - you get it with Kimi membership starting from the Plus plan at $19/month ($15/month with annual billing).
Status as of October 3, 2026
- Latest Kimi Code CLI is 2.1.1 from September 24, 2026 (we checked npm and GitHub), MIT license.
- Models: K3 (2.8 billion parameters, up to 1 million token context), K2.8 Preview under the old identifier
kimi-for-coding, and fast K2.7 Code HighSpeed. - On September 18, Moonshot resumed subscription sales suspended in July, with new Plus, Pro, Max and Ultra plans.
- On September 17, Kimi Code Desktop app launched for macOS and Windows.
- Free plan does not include Kimi Code.
What is Kimi Code
Kimi Code is a programming service included with Kimi membership. You use the same limit across four interfaces:
- Kimi Code CLI - terminal agent, single executable without Node.js. Has sub-agents, hooks, plugins, MCP servers and accepts video, e.g. screen recordings of errors.
- Kimi Code Desktop - application with built-in browser and plan mode.
- Kimi Code for VS Code - chat panel in the editor with change preview and undo.
- API key for third-party tools - subscribers create up to 5 keys in the Kimi Code console and connect K3 to Claude Code, OpenCode or Codex.
Zed and JetBrains connect to the CLI via ACP protocol (kimi acp). We wrote about the model and stock market reaction in Kimi K3 from Moonshot AI shakes up rankings.
How to Get Started Step by Step
- Purchase a plan starting from Plus on the Kimi membership page. The free plan does not provide access to Kimi Code.
- On Windows, first install Git for Windows - the CLI uses Git Bash as its shell.
- Install the CLI:
- macOS and Linux:
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash - Windows (PowerShell):
irm https://code.kimi.com/kimi-code/install.ps1 | iex - npm (Node.js from 22.19.0):
npm install -g @moonshot-ai/kimi-code
- macOS and Linux:
- Log in: in your project directory, type
kimi, then/login. Login works via device code - you open a link on any device and enter the code. - Give your first task, e.g. “describe the structure of this project”. Change models with
/model, check your limit with/usage, and update withkimi upgrade.
CLI Commands and Options - Tested by Us
On our Windows computer we have Kimi Code CLI 0.26.0 from July 16, 2026. On October 3, 2026 we only ran kimi --version and kimi --help, without any actual tasks. Key options:
| Command / flag | What it does |
|---|---|
kimi -p "command" |
Single task without interactive mode, output as text or --output-format stream-json |
kimi -m <model> |
Model selection for one run |
kimi --plan |
Start in plan mode: agent first proposes a plan, then modifies files |
kimi --auto / -y, --yolo |
Fewer confirmation prompts / automatic approval of everything |
kimi web, kimi server |
Browser interface and local REST + WebSocket server (defaults to 127.0.0.1 only) |
kimi acp |
ACP server for Zed and JetBrains |
kimi provider |
Adding other model providers, including from models.dev catalog |
In ~/.kimi-code/config.toml we see three models: K3, “K2.7 Coding” and “K2.7 Coding Highspeed”, each with 262,144 token context. The names are from July - kimi-for-coding has been running as K2.8 Preview on the server side since September 11. Our version is over a dozen releases behind current, so before testing we run kimi upgrade.
Models in Kimi Code
| Identifier | Model | Context | From plan |
|---|---|---|---|
k3 |
K3, most powerful; uses about 2x more limit than k3-256k | 256k, 1M from Pro | Plus |
k3-256k |
K3 with 256k context, no video input | 256k | Plus |
kimi-for-coding |
K2.8 Preview, according to Moonshot close to K3 | up to 1M | Plus |
kimi-for-coding-highspeed |
K2.7 Code HighSpeed: 5-6x faster, uses 3x more limit | 256k | Pro |
K3 launched on July 16, 2026 with open weights. Moonshot states it outperforms Opus 4.8 and GPT-5.5 - this is the manufacturer’s claim, not our measurement. You set reasoning level to low, high or max. After changing models, the documentation advises starting a new session, as old context falls out of cache and limit depletes faster.
How Much Does Kimi Code Cost
Prices from the kimi.com membership page, which we opened on October 3, 2026 (showing USD). Moonshot does not have a separate price list for other currencies.
| Plan | Monthly | With annual billing | Kimi Code |
|---|---|---|---|
| Free | $0 | - | No |
| Plus | $19 | $15 ($180/year) | Yes: K3 with 256k context, K2.8 Preview |
| Pro | $39 | $31 ($372/year) | Yes: K3 with 1M context, HighSpeed |
| Max | $99 | $79 ($948/year) | Yes, higher limit |
| Ultra | $199 | $159 ($1908/year) | Yes, highest limit |
In China the same plan names have different tiers: Go 49 CNY (without Kimi Code), Plus 99, Pro 199, Max 699 CNY monthly. Limits for new members:
- rolling 5-hour window and monthly pool, no weekly limit (old Moderato, Allegretto and other plans retain weekly window);
- CLI, VS Code, app and API keys share one limit;
- after exhaustion, Extra Usage kicks in: top-up from 25 CNY, billed close to API rates.
Moonshot does not publish message counts per window. With a Kimi platform API key you pay per token: K3 costs $3 per million input tokens and $15 per million output, K2.7 Code $0.95 / $4. We wrote about fluctuations in this rate in Kimi K3: price up 13%, and about the July sales suspension in Moonshot AI suspends new Kimi K3 subscriptions.
K3 in Claude Code and Other Tools
Yes, you can. The Kimi Code API supports OpenAI and Anthropic protocols. The official instruction for Claude Code involves entering the address https://api.kimi.ai/coding/, a key from the Kimi Code console, and a model like k3-256k into ~/.claude/settings.json. There are also instructions for OpenCode, Codex and Hermes Agent. Two notes:
- replacing the client identifier (User-Agent) is, according to Moonshot, a violation that can result in membership suspension;
- usage in third-party tools counts against the same limit as in the CLI.
Privacy: Where Your Code Goes
Kimi Code sends code and commands to Moonshot servers. What the documents reveal:
- Service for users outside China is provided by Singapore company NOVASCENT PRIVATE LIMITED. The privacy policy (updated August 13, 2026) states data may be stored outside the user’s country, and does not specify which country.
- The same policy permits use of user content, including files, for training models. We found no opt-out toggle in the policy text.
- The Chinese version (kimi.com, Beijing Moonshot) declares data storage in China. Documentation provides separate API addresses:
api.kimi.comfor China andapi.kimi.aifor the rest of the world. Our July CLI configuration hasapi.kimi.com. - CLI has anonymous telemetry enabled by default - you can disable it with
telemetry = falsein config.toml.
Our conclusion: fine for personal projects and open source code, but for client code only with owner consent. In the US there’s ongoing discussion about banning Chinese AI models, and general principles for safe work with agents are in our article about security gaps in vibe coding.
Our Test: Salary Calculator in Kimi Code
On October 3, 2026 we gave Kimi Code, Codex CLI and a local model the same task: a single-file gross-to-net salary calculator for 2026 with a 12-month table. We tested it in Polish, with Polish tax rules including tax thresholds and the 30x contribution cap. One rule: zero additional costs, no account creation and no re-login.
- Kimi Code CLI 0.26.0 (K3): in non-interactive mode
kimi -preturnedauth.login_requiredafter one second. The login token from July had expired, and re-login was outside our test rules, so we didn’t get a file. - The -p mode doesn’t combine with permission flags:
kimi -y -pandkimi --auto -pend with error “Cannot combine -prompt with…”. You need to remember this for automation. - For comparison, Codex CLI delivered the calculator in 2 minutes 28 seconds, accurate to the penny in all 10 checks.
Practical conclusion: if Kimi Code is to work in scripts or scheduled jobs, check login validity (kimi login) and update the CLI (kimi upgrade) before starting. Full test description, control values and screenshots are in The same salary calculator in Codex, Kimi Code and locally.
Kimi Code vs Competition
| Kimi Code | Claude Code | Codex | Antigravity | |
|---|---|---|---|---|
| Company | Moonshot AI (China) | Anthropic | OpenAI | |
| Cheapest access | Plus $19 | Pro $20 | ChatGPT Plus | free plan with limit |
| Main model | K3 (open weights) | Claude Opus 5.5 | GPT-6 Astra, GPT-6.1 Sol | Gemini, Claude 5.5 in paid plans |
| Open source client | Yes (MIT) | No | Yes (Apache-2.0) | No |
Broader overview: AI for programming. Common task for three agents: AI agent test: salary calculator.
Who Is It For
- For those seeking a powerful model cheaper - Plus at $15-19 gives you K3, also usable in Claude Code.
- For large repositories - from Pro plan K3 has 1 million token context.
- Not for sensitive corporate code without legal department approval.
Frequently Asked Questions
How much does Kimi Code cost?
From $19/month ($15/month with annual billing) on the Plus plan. Pro costs $39, Max $99, Ultra $199, and the free plan does not include Kimi Code.
Is Kimi Code free?
The client itself is free (MIT), but models require paid membership or an API key billed per token. You can also connect the CLI to other providers via kimi provider.
How to install Kimi Code on Windows?
Install Git for Windows, then in PowerShell run irm https://code.kimi.com/kimi-code/install.ps1 | iex, and in a new terminal type kimi and /login.
Can you use Kimi K3 in Claude Code?
Yes. Moonshot publishes official instructions: in ~/.claude/settings.json you set the address https://api.kimi.ai/coding/, a key from the Kimi Code console, and model k3 or k3-256k. Usage counts against your membership limit.
Is Kimi Code secure?
It has consent modes and blocks dangerous commands, but the privacy policy allows training models on user content and doesn’t specify data storage country. We don’t use it for corporate code without owner consent.
Kimi Code or Claude Code?
Claude Code gives you Anthropic models, Kimi Code is cheaper upfront and has an open client, and K3 can be run in Claude Code itself. Comparison on one task is in our salary calculator test.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources and Verification Date
- Kimi Code: documentation
- Kimi Code: membership and limits and model configuration
- Kimi Code: what’s new
- Kimi Code: instructions for Claude Code
- Kimi: membership pricing (USD, opened Oct 3, 2026)
- Kimi API Platform: model prices
- GitHub: MoonshotAI/kimi-code and npm: @moonshot-ai/kimi-code
- Kimi: privacy policy (updated Aug 13, 2026)
- KuCoin News: new Kimi plans after two-month pause (prices in CNY, Sept 18, 2026)
We quote prices in USD and CNY in original, without conversion. We verified CLI options on our own computer with kimi --help. Verified October 3, 2026.
More on this from QUANTSLANT
guides·ChatGPT Plus Codex Limits: Messages, 5-Hour Window & Weekly CapChatGPT Plus gives you 5-45 Codex messages per 5 hours in GPT-6 Astra, 15-160 in Sol, 350-3000 in Luna. Here's what affects usage and how to check your limits.

