Kimi K2 retirement: Moonshot closes the K2 API to new users — what changes, and what moving costs.

By AI Pro League · Published Oct 31, 2026

Kimi K2 retirement is a lifecycle story, not a price story. Our AI lab release tracker changelog records both K2 models in our database — Kimi K2.6 and Kimi K2.7 Code — moving from current to legacy status today, on the strength of Moonshot’s official API pricing page: the K2 series is closed to newly registered users from October 31, 2026, all K2 models are planned to retire by the end of December 2026, and existing users keep access until then. Moonshot fields 4 of the 59 models we track; the two K2 models are now half of its lineup here, and they join 8 legacy models in the database overall.

What the Kimi K2 retirement actually changes

• New API users cannot start on K2. From today, newly registered users no longer get the K2 series at all — a new project choosing Moonshot starts on K3 by default.
• Existing users have a deadline, not a cutoff. Access continues until the planned retirement at the end of December 2026 — roughly two months to migrate, on the pricing page’s own wording.
• Prices do not change. Both K2 models stay at the rates in the table below until they go — this is a withdrawal, not a price rise.
• The open weights stay available. Both K2 models in our database are open-weights releases, and the changelog records that the weights remain available after the hosted API retires — self-hosting is the one route that outlives the deadline.

The two K2 models in our database

Two details in that table deserve a second look. Kimi K2.6 still carries a League Score of 69.7 — retirement has not made it a bad model, only a departing one; its recorded component scores are on its model page. And Kimi K2.7 Code, the coding variant, has no League Score at all in our database — no recorded component scores — which is a coverage gap, not a judgement: an absent score contributes nothing and is never treated as a zero, the same rule our LMArena Elo explainer describes.

What moving to Kimi K3 costs — the worked example

The successors are already in the table: Kimi K3 and Kimi K3.1 Pro. Price the migration on a concrete workload — 10M input tokens and 2M output tokens, billed at each model’s published rates in our AI model pricing index:

The pattern is the story: Kimi K3.1 Pro costs $40.00 for the workload against $17.50 on Kimi K2.6 — about 2.3× the bill — for a League Score that is near-identical in our table. What the extra money buys is not a higher score today; it is a four-times-larger context window (1M against 262K tokens, as recorded) and a model line that is still open for business in January. Cached input softens the rise for repeated-context workloads — see prompt caching pricing explained for how those discounts work across the field.

If you build on Kimi K2, the checklist is short

1. New project? You are on K3 already — the K2 API door closed to new registrations today.
2. Existing K2 integration? Your deadline is the end of December 2026, per Moonshot’s pricing page — migrate on your own schedule inside that window, not in its last week.
3. Cost-sensitive workload? Re-price it before you move, with the table above as the template: the score barely moves, the bill does.
4. Cannot move? Both K2 models are open-weights and the weights remain available — self-hosting keeps the exact model, at your own infrastructure cost instead of an API bill.

Retirements like this are why our AI model rankings track status alongside scores: 8 of the 59 models in the database are now legacy, and a ranking that hid that would send new builders to a closed door. The Kimi story so far — K3 was also the League’s biggest riser in our October 26 ranking changes — is a lab replacing its own line in public, on a published timetable.