Moonshot Retires Kimi K2: New Signups Cut Off October 31 and Every K2 Model Goes Dark by End of December

Moonshot AI is sunsetting the model family that made Kimi a developer favorite. A notice posted on the company's API pricing documentation reads: "Dear Kimi Users, starting October 31, 2026, K2 series models will no longer be available to newly registered users; existing users are not affected. We plan to retire all K2 series models by the end of December 2026, please prepare your migration in advance."

Three models are affected: kimi-k2.6 ($0.95 input / $4.00 output per million tokens), kimi-k2.7-code (the same $0.95/$4.00 pricing, tuned for programming) and kimi-k2.7-code-highspeed ($1.90/$8.00). All K2 models carry a 262,144-token context window. Existing API users keep access until the December retirement date, but any new account created from October 31 onward cannot provision them.

The migration target is unambiguous: the documentation links directly to the Kimi K3.1 Pro quickstart guide. K3.1 Pro is priced at $2.00/$10.00 per million tokens — roughly double K2.6 on both sides of the ledger — with a highspeed variant at $4.00/$20.00 and a 1,048,576-token context window. Batch inference runs at 60 percent of standard pricing, which softens the effective increase for offline workloads.

The timing lands just after K3.1 Pro began appearing on third-party platforms, including ByteDance's Doubao Work agent, which added it alongside rival models earlier this week. Moonshot is effectively using the deprecation clock to push its installed base onto the new flagship — and onto pricier tokens — while the K3 generation competes against DeepSeek, Z.ai's GLM and Mistral's newly open-sourced trillion-parameter model on cost.

For developers, the notice is a forced re-test, not a drop-in swap. K2.7-code is embedded in a long tail of coding assistants, CI pipelines and agent frameworks where a $0.95/$4.00 model was doing the heavy lifting; doubling input costs changes routing decisions, especially for high-volume prefill workloads. The retirement also fits an industry-wide acceleration: OpenAI has scheduled GPT-4o and 4.1 snapshot removals for January 2027, and Google shut down its Veo 3.1 preview endpoints in late October.

Model churn is now an operating expense with a calendar attached. Teams running K2 in production have roughly eight weeks to benchmark K3.1 Pro against their own task sets, re-tune prompts written for K2's quirks, and decide whether the highspeed tier's latency is worth its premium. The labs have learned that defaults are destiny — and whoever controls the migration path controls the next generation of usage.