Kling AI has set September 15, 2026 as the retirement date for a large block of legacy models and APIs. Anyone integrating with the platform has a hard deadline.

What is being retired

Video models: Kling 1.0, Kling 1.5, Kling 1.6, Kling 2.0 Master, Kling 2.1 and Kling 2.1 Master.

Image models: Kling Image 1.0, 1.5, 2.0 and 2.0 New.

Also going: the Virtual Try-On API and 119 video effect templates from the Video Effects Center.

Kling recommends Kling Image 3.0 and 3.0 Omni as alternatives. For virtual try-on it says a next-generation experience is in development, with no date given — so that capability has a gap rather than a migration path.

What survives

Content generated before the retirement date is unaffected and remains available. The shutdown removes the ability to call the old models, not the output already produced with them.

The current line is Kling 3.0 and 3.0 Omni, Kling 3.0 Turbo, Kling Image 3.0 and Kling Motion.

Why it matters

Six video model versions disappearing at once is unusually aggressive. Most vendors stagger deprecations; Kling is clearing everything before 3.0 in a single date.

For anyone with production workflows pinned to a specific version, this is real migration work, and the deadline is days away rather than months. Generated output does not change, but prompts tuned against 2.1 will not behave identically on 3.0 — model upgrades shift style and timing even when they improve quality.

What to do now

Check which model IDs your integration actually calls. If any of the retiring versions appear, test the 3.0 equivalents against your own reference prompts before the 15th rather than after. Virtual Try-On users need a different plan entirely, since there is no replacement available yet.

The wider pattern

Video generation vendors are consolidating around single current lines rather than maintaining back catalogues, and the reason is cost: serving six versions of a video model means six sets of weights on expensive hardware. Text model providers can afford long legacy tails; video providers apparently cannot.

For anyone building on generative video, that argues for treating the model version as configuration rather than as a fixed assumption, and for keeping a reference set of prompts you can re-run whenever a vendor forces a move.