SlopTV

news

Google sets September 30 as the deprecation date for the gemini-omni-flash-preview endpoint

The Gemini API release notes, carrying a September 17 update stamp, retire the omni preview route and point developers at a new Gemini Omni Flash model page, while listing Veo 3.1 Lite Preview as the cheap video option.

By Priya Shenoy ยท

Google's Gemini API release notes state that the existing gemini-omni-flash-preview endpoint will be deprecated on September 30, 2026, and direct developers to the Gemini Omni Flash model page and omni guide instead. The same notes list Veo 3.1 Lite Preview, veo-3.1-lite-generate-preview, as Google's most cost-efficient video generation model.

Anyone still calling Google's omni model through the preview endpoint has until the end of this month. The Gemini API release notes state that the existing gemini-omni-flash-preview endpoint will be deprecated on September 30, 2026, with developers pointed to the Gemini Omni Flash model page and the omni guide as the route forward. The page carries a last-updated stamp of September 17, 2026 UTC, which puts the notice roughly ten days ahead of the cutoff.

The notes give a date and a destination, not a mapping

What the release notes provide is narrow: an endpoint name, a deprecation date, and two documentation links. There is no parameter-level migration table in the changelog entry, no statement that request bodies are compatible between the preview route and whatever replaces it, and no grace period described beyond the September 30 date itself. For a preview endpoint that is the expected pattern, since preview strings in the Gemini API have never carried stability guarantees, but it does mean the work of checking whether a production call still returns the same shape falls entirely on the caller.

The Gemini Omni line matters here because it is Google's multimodal output surface rather than a Veo-branded video model, so teams that adopted it were often doing so for mixed text, image, audio and video work in one call. That is exactly the kind of integration where a swapped model string is the cheapest part of the migration and the output-handling code is the expensive part.

Veo 3.1 Lite Preview sits in the same release notes as the low-cost video route

The same changelog records the launch of Veo 3.1 Lite Preview under the model ID veo-3.1-lite-generate-preview, which Google describes as its "most cost-efficient video generation model" and positions for rapid iteration and high-volume applications. Google does not state a per-second or per-generation figure in the release notes entry, so the cost claim sits as a relative statement against the rest of the Veo 3.1 family rather than a number anyone can check against a competitor.

Worth reading alongside it: the same release notes also carry an earlier deprecation announcement covering a set of image generation models that were shut down on August 17, 2026. Two image-model shutdowns and one omni endpoint deprecation inside a single changelog is a reasonable picture of how fast the generative media side of the Gemini API is turning over compared with the text models, where version numbers keep climbing but endpoints tend to survive longer.

Two video API deadlines land six days apart

September 30 is not the only date on the calendar for video API consumers this month. OpenAI's Sora API is scheduled to sunset on September 24, 2026, which we covered when Disney's character licensing deal ran straight into the shutdown of the API you would build on. A team that picked both a US frontier video API and Google's omni preview route in the first half of 2026 is therefore doing two migrations in the same week, for two different reasons: one because the product line is being retired outright, the other because a preview string is being rotated.

The practical read is that preview endpoints in generative media are now short-lived enough that pinning one in production is a scheduling decision, not a technical one. Google's notes do not say the omni capability is going away, only that this particular address for it is. That distinction is the whole difference between a find-and-replace and a rebuild, and the release notes leave it to the reader to work out which one applies to their code.

Sources: Google Gemini API release notes.


PS

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.