Kimi K2 retirement: new users cut off, all K2 gone by December

The Kimi K2 retirement has a date now. As of October 31, 2026, Moonshot AI no longer offers K2 series models to newly registered API users, and it plans to retire all K2 series models by the end of December 2026. If your agent or app calls a K2 model on the Kimi API, you have about eight weeks to move it. If you planned to spin up a new account to run K2 for a client, that door already closed.

What actually happened

The notice sits at the top of Moonshot's own model list and pricing page. It says three things:

• Since October 31, 2026: K2 series models are not available to newly registered users. Existing users are not affected yet.
• By end of December 2026: Moonshot plans to retire all K2 series models and tells users to prepare the migration in advance.
• The replacement: Moonshot points everyone to kimi-k3.1-pro, which launched October 23. Moonshot lists a 1,048,576-token context window for all K3 series models, and says batch inference on kimi-k3.1-pro is billed at 60% of the standard price.

This is not Moonshot's first cleanup this year. The same page says the older kimi-k2 preview and thinking models were discontinued on May 25, 2026, and the moonshot-v1 series on August 31, 2026.

One gap: the notice says "K2 series" without a full list of model IDs. The current list still shows kimi-k2.6, kimi-k2.7-code, and kimi-k2.7-code-highspeed. Third-party migration guides say those are covered, but Moonshot's page doesn't spell it out. Assume every model ID that starts with kimi-k2 is on the clock.

Why the Kimi K2 retirement matters for your business

Cheap open-weight models are a good deal right up until the hosted version goes away. A model swap is not a string change. Prompts tuned for K2 can break on K3.1 Pro: different tool-call habits, different verbosity, different reasoning output, a different bill.

1. Grep for the model ID. Search code, env files, and agent configs for kimi-k2. Hardcoded IDs in five places is how a deprecation turns into an outage.
2. Run your evals on K3.1 Pro now. Use your real tasks, not the launch benchmarks. Compare output quality, tool calls, and cost per completed task.
3. Check your price. We couldn't confirm K3.1 Pro rates on Moonshot's pricing page today. Pull the number from your console before you budget the switch.
4. Keep a fallback route. Route through one config value so the next retirement is a one-line change, not a sprint.
5. Own the weights if you need them. K2 weights are open. If a workflow must stay on K2, self-hosting is an option. The hosted API is not.

Sources: Kimi API Platform: Model List, Kimi API Platform: Pricing.