Moonshot to retire its Kimi K2 API models by the end of December

Moonshot to retire its Kimi K2 API models by the end of December

25 October 2026: Moonshot AI will stop offering its K2 series models to anyone who registers on the Kimi API platform from 31 October 2026, and plans to retire the whole series by the end of December. Teams already calling kimi-k2.6 or kimi-k2.7-code keep access for now, but the notice tells them to "prepare your migration in advance", which in practice means moving to the K3 line before the year closes.

Not yet independently verified. This rests on the Kimi API platform's own pricing and model list pages, read on 25 October; neither page dates the notice, and no exact retirement day has been given beyond "the end of December 2026". We will update this when it can be confirmed, and remove this note.

The same notice now sits at the top of two pages of the platform's documentation: the pricing explanation and the model list. It reads: "starting October 31, 2026, K2 series models will no longer be available to newly registered users; existing users are not affected." Neither page dates the notice, and Moonshot has not named a day for the retirement itself.

The three models still on the K2 list

The model list names three K2 models that are still live and therefore covered by the notice:

• kimi-k2.7-code, which Moonshot describes as its dedicated coding model, with a 256k context
• kimi-k2.7-code-highspeed, the faster variant, which the page puts at about 180 tokens a second and up to 260 in short-context use
• kimi-k2.6, which takes text and image input, has thinking and non-thinking modes, and also runs with a 256k context

The pricing page gives the exact K2 window as 262,144 tokens. Every older K2 model is already gone. The original kimi-k2 series, including kimi-k2-0711-preview, kimi-k2-0905-preview, kimi-k2-turbo-preview and the two thinking variants, was discontinued on 25 May 2026; kimi-k2.5 and the whole moonshot-v1 family followed on 31 August. For each of them the page points users to kimi-k3.1-pro.

Where the K3 line differs

The replacement models are kimi-k3, which Moonshot calls its previous-generation flagship and an open-source model of 2.8 trillion parameters, and kimi-k3.1-pro with its high-speed variant. All K3 models take a context of 1,048,576 tokens, four times the K2 window.

The billing also works differently, and this matters more than the window for anyone budgeting a migration. The platform bills inference pay as you go, with no subscription. For K3 models, the pricing page says, "cache writes are billed separately by TTL (5min / 1h)", while cached input is charged at the cache-hit price with no further write charge. Context caching switches on automatically for repeated request prefixes with a five-minute default. An application that sends long, repeated system prompts will therefore see a new cache-write line on its bill after the move, and the choice between the five-minute and one-hour tiers changes what it pays. The price tables themselves did not render in the copy of the page we read, so we do not quote per-token rates here.

Who has to act before the end of December

Two groups are exposed. Anyone planning to start a new project on a K2 model, for example to keep costs down on a coding agent, has until 30 October to register; after that the K2 models will not be offered to the new account at all. Existing integrations keep working, but on a clock: when the series retires, the K2 models join kimi-k2.5 on the list of discontinued models that are "no longer maintained or supported". Code that hard-wires a model name, or a gateway that maps an internal alias to kimi-k2.7-code, needs the change made and tested on K3 before then, including a check of how the larger context and the caching tiers affect cost.

Our AI model category lists 135 models, among them Kimi K3 and Kimi K3.1 Pro, the two models Moonshot names as the way forward. The retirement day itself is still open: Moonshot has said only "the end of December 2026".

Company profile on TrustList: Moonshot AI

Sources

• Kimi API Platform: Model Inference Pricing Explanation (read 2026-10-25), 25 October 2026
• Kimi API Platform: Model List (read 2026-10-25), 25 October 2026