Moonshot closes Kimi K2 to new API sign-ups

If you build on Kimi K2, you have about two months. And the move to K3.1 Pro is not a model-name swap.

Moonshot AI's pricing page now says that "starting October 31, 2026, K2 series models will no longer be available to newly registered users." Existing users are not affected for now. Moonshot says it plans "to retire all K2 series models by the end of December 2026," and asks developers to prepare their migration in advance.

The door closed on Saturday

New sign-ups lost K2 on Oct. 31. Existing accounts keep it until the retirement Moonshot plans for the end of December.

The notice points developers to the Kimi K3.1 Pro quickstart. That page describes K3.1 Pro as a "3T-class open-weight model" with a 1 million token context window. A faster variant, kimi-k3.1-pro-highspeed, runs two to three times faster, Moonshot says.

TopFive covered the K3.1 Pro launch, when Moonshot said its weights would follow by Nov. 4.

Why it matters: the defaults change under you

A retirement notice reads like housekeeping. This one changes how your app behaves.

Moonshot's quickstart lists what is different on K3.1 Pro:

• Thinking is always on, and it can't be turned off.
• Reasoning effort accepts low, high or max, and the default is max. Setting low is the only way to cut reasoning time.
• Temperature is fixed at 1.0 and top_p at 0.95, along with fixed n and penalty settings. Moonshot says to leave them out of requests.
• In multi-turn and tool-call conversations, you return the complete assistant message unchanged, not just its content.
• Vision input doesn't accept public image URLs. You send base64 or an uploaded file ID.

Read that list as a developer. A K2 call that passed a low temperature for consistent output now has no such knob. A call that returned only the assistant's text in a tool loop may break. And a default of max effort means more thinking per request than you may have planned for.

That is latency and spend you didn't choose.

Inventory first, then test

Start with an inventory. Find every place your code calls a K2 model, and note any request that sets temperature, top_p or penalties.

Then run your own evals on K3.1 Pro at low and at max effort. Compare quality, response time and token use against K2 on the same tasks. Moonshot's pricing page says usage is billed pay-as-you-go per million tokens, with cached input billed at a cache-hit price, so token counts are the number to watch.

If you'd rather own the model than rent it, the open weights Moonshot says are due are the other path. A 3T-class model is a lot of hardware to run yourself, so price that before you count on it.

And if you are a new developer looking at Kimi for the first time, K2 is already gone for you. Start on K3.1 Pro.

What to watch

Watch for the K3.1 Pro weights by Nov. 4, and for any firmer date on K2's end in December.

Then set a reminder for early December. Don't find out from an error log.