EVERYDAY
Kimi K2.6
Fast, low-cost chat and Q&A. The only current model where thinking can be switched off.
- Thinks
- On or off
- API
- $0.95 / $4
- Context
- 262K
Kimi offers several models, and picking the right one saves time and credits. This page explains each current model in plain words, what "reasoning effort" means, and which model fits which job, whether you use the Kimi app or build with the API. It also lists the older models that have been retired.
Match the model to the job. These pairings follow Kimi's own help centre.
Kimi K2.6. Fast, and the cheapest on credits and on the API.
Kimi K3. The strongest model, and it makes real files.
Kimi K2.7 Code, through Kimi Code or the API.
A model is the "brain" behind the assistant. These are the ones Kimi offers today.
Fast, low-cost chat and Q&A. The only current model where thinking can be switched off.
A specialist for programming, built on K2.6. Also comes in a faster HighSpeed version.
The most capable model. Makes real files such as slides and spreadsheets, and reads images and video.
Kimi K3 working as a team of up to 300 agents, for huge searches and batch jobs. In the Kimi app.
API prices are per million tokens, input / output, before tax, from Moonshot's price page. Context is the API figure in tokens. K3 Cluster details are from Kimi's help centre.
In the Kimi app, you choose a model and how hard it should think. More effort usually means better answers on hard problems, but it uses more tokens and so more credits.
The fastest and cheapest. Kimi's tip for saving credits. On K2.6, K3 and K3 Cluster.
Stronger reasoning for complex analysis and difficult problems. On all three.
The deepest thinking. Only on K3 and K3 Cluster.
App names from Kimi's help centre. On the API, the equivalent settings differ by model: Kimi K3 uses reasoning_effort with low, high or max, and Kimi K2.6 switches thinking on or off. See Kimi K3 API and Kimi K2.6 API.
| K2.6 | K3 | K3 Cluster | |
|---|---|---|---|
| Best for | Fast chat and Q&A | Documents, slides, sheets and multi-step tasks | Large searches, batch jobs, very long writing |
| Effort levels | Standard, Advanced | Standard, Advanced, Extreme | Standard, Advanced, Extreme |
| Makes files | Text only, such as outlines | Yes: .docx, .pptx, .xlsx, .pdf | Not stated |
| Images, video, audio | Not stated | Yes, through plugins | Not stated |
| Reads at once | About 128K tokens a turn | Up to 1 million, on the top tier | Not stated |
| Credits | Uses credits | Uses credits | Uses credits |
On a small screen, swipe the table sideways to see every column.
From Kimi's "How to Choose" help article. It does not give per-model credit rates or say which plans include which model, apart from the 1-million-token window. "Not stated" means the article does not say. More: Kimi AI pricing.
| Model name | Input | Output | Context | Thinking |
|---|---|---|---|---|
kimi-k3 | $3.00 | $15.00 | 1,048,576 | Always, low / high / max |
kimi-k2.7-code | $0.95 | $4.00 | 262,144 | Always |
kimi-k2.7-code-highspeed | $1.90 | $8.00 | 262,144 | Always |
kimi-k2.6 | $0.95 | $4.00 | 262,144 | On or off |
On a small screen, swipe the table sideways to see every column.
kimi-k2.6. Use kimi-k2.7-code for coding agents. Send only the hard or very long requests to kimi-k3.Prices per million tokens, before tax, from Moonshot's price page. More: Kimi API pricing · Kimi API setup · K2.7 Code HighSpeed
Kimi also has several products. A product is where you work; a model is the brain inside it. Kimi's help centre describes them like this.
The website and apps. Chat, search, Deep Research, slides, sites and sheets.
A desktop app for Mac and Windows that works with files on your computer.
For developers, in the terminal and VS Code.
An always-on agent in the cloud, with chat-app connections.
Kimi says chat, agents, Deep Research and Kimi Code share one credit pool. More: all Kimi features · Kimi Code · Kimi Claw
Comparisons: Kimi K2 vs K2.5 · Kimi vs other AI tools
These names no longer work on Moonshot's API. Moonshot's advice is to move to kimi-k3.
| Model | Retired | Consider instead |
|---|---|---|
kimi-k2.5 | 31 August 2026 | kimi-k2.6 or kimi-k3 |
kimi-k2-*, including thinking and turbo | 25 May 2026 | kimi-k2.6, kimi-k2.7-code or kimi-k3 |
On a small screen, swipe the table sideways to see every column.
Dates from Moonshot's models page. The "consider instead" suggestions beyond kimi-k3 are ours. Our home page also lists older names, such as moonshot-v1, as retired; we did not re-check those today. The open weights of K2 and K2.5 can still be downloaded.