How to migrate off Kimi K2.5 before it shuts down

Why is Kimi K2.5 being retired?

Moonshot AI is consolidating the Kimi lineup around K3, released on 16 July 2026. K2.5 shipped in January 2026 and has been superseded twice since — by K2.6 in April and K2.7 Code shortly after. The platform documentation now lists K2.5 alongside the moonshot-v1 series as unavailable to newly registered users, with a full platform sunset on 31 August 2026.

This is the second retirement in four months. The entire kimi-k2 series, including kimi-k2-thinking and the turbo variants, was discontinued on 25 May 2026 and those endpoints no longer respond. The pattern is worth internalising: Moonshot retires models on a timeline measured in months, not years, and it does not appear to extend deadlines.

Which model should replace Kimi K2.5?

It depends on which property of K2.5 you actually depend on. Three of the current models are reasonable targets, and only one of them is a like-for-like swap.

If you rely on	Move to	Why
General-purpose chat and agents	Kimi K2.6	Same 262,144-token window, same multimodal inputs, no request-format change
Coding workloads	Kimi K2.7 Code	Coding-specialised, priced identically to K2.6 on cache-miss input and output
Context beyond 262,144 tokens	Kimi K3	The only Kimi model with a 1,048,576-token window
Lowest possible latency	Kimi K2.7 Code HighSpeed	~180 tokens/second, but exactly double the price

For most people reading this, the answer is Kimi K2.6. It is the direct successor: identical context window, identical modality support, and no changes to the request schema.

What does the migration actually involve?

In the common case, one line. Both models are served from the same Chat Completions endpoint with the same parameters, so the change is the model identifier:

  response = client.chat.completions.create(
-     model="kimi-k2.5",
+     model="kimi-k2.6",
      messages=messages,
  )

What does not transfer automatically is prompt behaviour. K2.6 is a different model, and prompts tuned against K2.5 — particularly ones that depend on a specific refusal style, output format or reasoning verbosity — can behave differently. Before you ship:

• Run your longest realistic prompt through the new model and confirm the response still fits inside the 262,144-token window.
• Test every image and video input path, since multimodal behaviour changes between model generations even when the API surface does not.
• Re-run whatever evaluation set you have. If you do not have one, a migration forced by a vendor deadline is an excellent reason to build a small one.

How much more will Kimi K2.6 cost?

More, on the headline rates — but the size of the increase depends heavily on your cache hit rate, which is the part most cost comparisons skip.

	Kimi K2.5	Kimi K2.6	Change
Input, cache hit	$0.10	$0.16	+60%
Input, cache miss	$0.60	$0.95	+58%
Output	$3.00	$4.00	+33%

All figures are USD per million tokens, excluding tax, from Moonshot’s own pricing pages as verified on 6 August 2026.

The output increase is the smaller one at 33%, and output usually dominates conversational workloads. Agent workloads that resend a large fixed system prompt see a different picture, because their input volume is large and mostly cached.

Work out your own number rather than trusting a percentage — the API cost calculator takes your token volumes and cache hit rate and shows every model side by side.

What happens if you miss the deadline?

Your requests fail. There is no grace period documented, and the precedent from the kimi-k2 retirement in May 2026 is that the endpoints simply stop responding.

Two practical mitigations if you are close to the date:

1. Fail over rather than fail. Wrap the model identifier in configuration rather than hardcoding it, so switching is a deploy and not a code change. This is worth doing regardless — there will be another retirement.
2. Do not rely on being warned. Moonshot publishes deprecations in its platform changelog, not by email to every API key holder. If you are running Kimi in production, that changelog belongs in your monitoring.

Is self-hosting K2.5 a real option?

Technically yes. Moonshot published open weights for K2.5, and the platform sunset does not withdraw them. Practically, this is a large commitment: K2.5 is a 1-trillion-parameter model, and serving it at useful latency requires substantial GPU capacity that you now own and operate.

Self-hosting makes sense if you have a hard requirement that the hosted API cannot meet — data residency, air-gapped deployment, or a genuinely enormous and predictable request volume. It is almost never the cheaper answer to “my vendor is retiring a model”, and treating it as one is how teams end up maintaining inference infrastructure they never wanted.