AI video's preview trap keeps retiring models mid-project

On October 22, 2026, three Veo video models on Google's Gemini API reached the shutdown date that Google's own deprecations page had given them. The replacement the page names is gemini-omni-1.1-flash.

The date is the news. The table it comes from is public, and it already lists the dates after this one. Google's page is one of a small set of maker-run lists that publish, in plain text, the exact day each model stops answering, and those lists are busier than a creator building on top of them might hope.

What Google switched off on October 22

Three rows in the Veo section of the deprecations page carry the same date and the same recommendation:

• veo-3.1-generate-preview, released October 15, 2025, shutdown date October 22, 2026, recommended replacement gemini-omni-1.1-flash.
• veo-3.1-fast-generate-preview, released October 15, 2025, shutdown date October 22, 2026, recommended replacement gemini-omni-1.1-flash.
• veo-3.1-lite-generate-preview, released March 31, 2026, shutdown date October 22, 2026, recommended replacement gemini-omni-1.1-flash.

A fourth endpoint, gemini-omni-flash-preview, released June 30, 2026, carries the same October 22 date and the same replacement. Google's video generation guide points the same migration at Gemini Omni Flash, gemini-omni-1.1-flash, or the generally available Veo models on the Gemini Enterprise Agent Platform.

The page explains its own terms. A deprecation "is the announcement that we no longer provide support for a model, and that it will be 'shut down' in the near future," it reads, and once a model is shutdown, "it is completely turned off, and the endpoint is no longer available."

Veo models have been reaching that point for a year. The same table shows veo-3.0-generate-001 and veo-3.0-fast-generate-001, both released September 9, 2025, and veo-2.0-generate-001, released April 9, 2025, all shutting down on June 30, 2026. Before them, veo-3.0-generate-preview and veo-3.0-fast-generate-preview shut down on November 12, 2025. Each row names a successor, and for the Veo 3.0 and Veo 2.0 retirements the successor named was itself a Veo 3.1 preview build, which now carries the October 22, 2026 date.

The model now receiving all of them, gemini-omni-1.1-flash, was released August 27, 2026. Its row carries no shutdown date announced.

Google's page is careful about how it words those dates: it tracks the "announced earliest shutdown dates," which leaves a model's last day open to change. The table is not only about video either. deep-research-pro-preview-12-2025 carries a shutdown date of October 23, 2026, and five speech, live-audio and text-to-speech previews are dated November 17, 2026.

The Sora shutdown is the extreme end of the same pattern

OpenAI's deprecation page files the Sora 2 shutdown under past deprecations. "On March 24th, 2026, we notified developers using the Videos API and Sora 2 video generation model aliases and snapshots of their deprecation and removal from the API on September 24, 2026," the entry reads.

Six rows share that one date: the Videos API itself, the model aliases sora-2 and sora-2-pro, and the dated snapshots sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. The recommended-replacement column is blank for all six. Other retirements in the same table name a successor model. This one does not.

OpenAI's own help page splits the shutdown in two: the Sora web and app experiences were discontinued on April 26, 2026, and the API followed on September 24, 2026. Content export runs through a sunset page, and the company says it will permanently delete data tied to the product once any final export window passes. The Decoder reported the two-stage plan when it landed.

The pattern is not specific to one maker, and the next two examples show how differently it can be handled.

Kling retired a stack of models on one date

Kling did not retire one model. Integrators that carry Kuaishou's Kling models recorded a single date for a whole block of them.

ComfyUI's partner-node documentation states that "the following legacy Kling models and workflow templates are scheduled for retirement on September 15, 2026: Kling V1.5, V1.6, V2.1, V2.1 Master, and the Kolors Virtual Try-On API." The Krea changelog lists "Kling 1.0, 1.5, 1.6, 2.0 Master, and 2.1" and says their API endpoints "are deprecated and sunset on September 15, 2026." Krea adds the detail a caller actually meets: until that date the endpoints returned Deprecation and Sunset headers, and after it they "return 410 Gone."

Leonardo's API documentation carries the same date for one of them, timed to the minute: "Kling 2.1 Pro is being retired on Tuesday, September 15, 2026 at 9:30 AM AEST (UTC+10)," with a recommendation to move to Kling Video 3.0 Pro before then.

Kling's own API updates page records the other side of the same day. Under 09/15/2026 it lists "Virtual Try-On Version 3.0 Officially Released," and notes that "V1/V1.5 model parameters have been adapted on the engineering side, allowing users to seamlessly switch to the 3.0 pipeline."

A quieter version of the same signal sits on Luma's own information page, which tells readers to refer to the current video model, Ray3.2, and warns that "Dream Machine or Ray2, which are deprecated," should not be conflated with it.

Why the treadmill runs

The engine is the preview model, and OpenAI states the reason on its own page.

Its deprecation policy sets three notice periods. Generally available models get at least six months. Specialized variants of those get at least three. Preview models, "identified by preview in the model name, may be retired with much shorter notice, such as 2 weeks." OpenAI adds a plain warning: "We don't recommend using preview models for business-critical production workloads unless you can migrate on short notice."

Read the Gemini dates against that. gemini-omni-flash-preview was released June 30, 2026 and was dated to shut down October 22, 2026, about 16 weeks later. veo-3.1-lite-generate-preview got a little under seven months. The two older Veo 3.1 previews, released October 15, 2025, lasted about twelve months.

The gaps are not uniform. A creator who plans around a model's quality is planning around a moving target, because the version most likely to be retired early is the one that just arrived.

What actually breaks when an endpoint retires

This part is analysis, not a claim about any one maker, but the mechanics repeat across all of them.

The first break is a pinned string. A script, a notebook, a config file or a saved request that names a retired model fails on the next call. Krea's note that retired Kling endpoints "return 410 Gone" is exactly the failure a caller meets: the code did not change, and the model stopped existing.

The second break is an automation flow. A workflow, template or scheduled job that carries a model name breaks when the node behind it is pulled, and that can happen at load time rather than at render time. ComfyUI's own changelog records both halves of this, from "retired Sora video nodes are gone from ComfyUI" to a run of "retired partner models" removed in a single release.

The third break is anything stored on the maker's side. When an API and its storage go away together, finished output sitting in the vendor's account goes with them unless it was pulled down first, which is why the Sora notice puts an export deadline ahead of a deletion.

A creator can defend against the first two without knowing which models are next.

How to make the next shutdown boring

This is the practical half, and it is analysis.

Keep your own copies. Prompts, seeds, source images and finished renders belong in folders you control, not in a vendor's project view. The Sora export mechanism exists because the alternative is losing the work.

Pin the model string in one place. If every flow that calls a model points at a single named constant, a retirement becomes one edit instead of a search across a project.

Read preview as a countdown. A model ID with the word preview in it is a bet against a shorter clock, and OpenAI's own page says so in as many words. Where a generally available version or a dated snapshot of the same capability exists, it carries a longer promise.

Watch the makers' tables, not the launches. The shutdown dates live on the deprecation pages, which is where the Kling date surfaced for ComfyUI, Krea and Leonardo before the endpoints stopped answering.

Google's table already holds the next entries. November 17, 2026 carries five preview models in the speech, live-audio and text-to-speech sections, each with a named replacement. The October 22 row is now the recent past, and the table carries two more dates after it: October 23, 2026, for deep-research-pro-preview-12-2025, and November 17, 2026.

Google keeps its schedule at ai.google.dev/gemini-api/docs/deprecations and OpenAI at developers.openai.com/api/docs/deprecations; both are public pages worth bookmarking next to whichever maker your own pipeline actually calls.