Kimi Closes K2 to New Users – Move to K3.1 Pro by December
In short
- Moonshot stopped giving the Kimi K2 series to newly registered API users on October 31. K2 is planned to be retired by the end of December 2026.
- The replacement is kimi-k3.1-pro with a 1 million token context. Full open weights are promised by November 4.
- The money angle: K3.1 Pro is reported at $2 input and $10 output per million tokens, double GPT-6.1 Terra’s $1 and $5. The open weights are the real value, not the price.
If your app runs on Kimi K2, you have two months. Moonshot just closed K2 to new users, and the retirement is planned for the end of December. Existing users keep access for now. Every new project starts on K3.1 Pro.
What happened On October 31, Moonshot’s API docs changed: K2 series models are no longer available to newly registered users. Existing users are not affected yet. All K2 models are planned for retirement by the end of December 2026, and the docs point everyone to kimi-k3.1-pro.
K3.1 Pro itself is not new. It reached the API on October 23. It has a 1 million token context window, and a kimi-k3.1-pro-highspeed variant that Moonshot says runs 2 to 3 times faster. To use the flagship you top up at least $1, and the total you top up sets your rate limits. Moonshot promises the full weights by November 4.
What does it cost? The quickstart page I checked does not list rates. Third-party reports citing Simon Willison give $2 per million input tokens and $10 per million output. Check Moonshot’s pricing page before you budget.
How to use it – Already on K2? Switch the model name to kimi-k3.1-pro in a test setup now. Compare the answers on your 20 most common prompts. Do not wait for December. – Starting a new project? You cannot pick K2 anyway. Start on K3.1 Pro, or the highspeed variant if response time matters. – Working with long documents? One million tokens fits a whole contract bundle or a codebase in a single request. – Want to self-host? Wait for November 4. Until the weights are out, nothing can be self-hosted, and the hardware needs are unknown.
What it means Labs now retire models within months. A model name on an API is a rental, not something you own. That is why the open weights matter: if Moonshot ever switches a model off, weights you already downloaded keep working.
Do the math: a workflow uses 10 million input and 2 million output tokens a month. On Kimi K3.1 Pro at the reported $2 and $10, that is 10 x 2 + 2 x 10 = $40 a month (about €35), $480 a year. On GPT-6.1 Terra at $1 and $5, it is $20 a month (about €17), $240 a year. So Kimi costs twice as much per token. Per-token price is not the whole story, because the models differ in quality and in how many tokens they burn per task. But you are paying for the 1 million context and the open weights, not for savings.
Ivan’s Take
Kimi K3.1 Pro is not the cheap option. Terra costs half as much per token. I would move off K2 only when I have to, and I would wait for November 4 to see if the weights change the math. Open weights are the one thing a closed API cannot give you: a model nobody can switch off.
Source: Kimi
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
