Google Is Retiring Every Veo Endpoint: Inside the October 22 Video Shake-Up

A deadline with no escape hatch

Model deprecations usually come with a soft landing. You get a warning, you get a window, and if you miss the window you can pin the old version and buy yourself a quarter to migrate properly. What Google put on its Gemini API deprecation page for October 22, 2026 is the other kind of deadline. There is nothing to pin to.

Every Veo video endpoint on the Gemini API is scheduled to shut down on the same day: veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview. The earlier Veo generations are already gone. Veo 3.0 and 2.0 were retired on June 30, 2026. After October 22, the number of Veo models callable on the Gemini API is zero.

The reason becomes obvious when you read the suffixes. Every one of those identifiers ends in -preview. Google has served Veo on this API for roughly eighteen months and never graduated it to general availability. The standard advice, pin the last stable release and move on your own schedule, has no target because there was never a stable release to pin.

What replaces it, and why the math is strange

Google names one replacement for all three Veo endpoints: gemini-omni-1.1-flash, which went GA on August 27, 2026. It is not Veo 4. It is not a new Veo at all. It is a different model family, served through a different interface, and it changes the shape of the bill.

Veo 3.1 billed by the second of finished video. The rates were tiered by resolution: the standard endpoint at $0.40 per second at 720p and 1080p, rising to $0.60 at 4K; the Fast endpoint at $0.10 to $0.30; the Lite endpoint at $0.05 to $0.08. Gemini Omni bills per output token instead. Converting between a per-second and a per-token model is the part no vendor page does, because it requires committing to an assumption about how many tokens a second of video costs.

When someone does the arithmetic, the result depends entirely on which Veo variant you were on. For teams on the standard Veo 3.1 endpoint, the move comes out as a substantial price cut, on the order of three quarters off. For teams on Fast, it is roughly a wash. For teams on Lite, the cheapest video-with-audio endpoint Google has ever sold, the bill roughly doubles.

The pattern is worth sitting with. The cheapest option gets punished the hardest. Anyone who chose Lite specifically because it was cheap, which is the entire reason to choose a Lite tier, is being handed a migration and a bigger number at the same time. Vendors rarely plan this outcome deliberately. It falls out of a simpler rule: the endpoints a company is least invested in are the ones it is most willing to retire, and the ones it is least invested in are usually the cheap ones.

The short life of a cheap endpoint

One row in the deprecation table deserves its own paragraph. veo-3.1-lite-generate-preview shipped on March 31, 2026 and is scheduled to go dark on October 22, 2026. That is a 205-day life. Less than seven months from launch to shutdown, on a model that was positioned as the accessible entry point.

This is the clearest illustration of a risk that has been building all year: in AI infrastructure, adopting the cheap tier is not the low-risk choice. The cheap tier is often the experimental one, and experimental things get retired when the experiment ends. The premium tier, expensive as it is, sometimes has more staying power because the vendor has committed more to it.

Read the fine print on the dates

There is one mitigating detail, and it cuts both ways. Google's own page describes the shutdown dates as "the earliest possible dates on which a model might be retired." A slip is possible. Anyone building a migration plan on the hope of a slip is building on sand, but anyone treating October 22 as a hard wall should know the vendor has left itself room.

What is not in question is the direction. The Gemini API is consolidating its video story onto a single GA model in the Omni family. The voice endpoints are on the same path, with text-to-speech and live previews flagged for shutdown no earlier than November 17, 2026, and the replacement speech models carrying a scheduled price change on January 1, 2027. A team that migrates its voice pipeline in November gets one price for about six weeks and a different one after that.

Who gets hurt

The damage from a migration like this is not spread evenly. Large teams with a platform group and a vendor relationship absorb it as a sprint. Small teams that built a product on the cheap endpoint absorb it as a crisis, because their unit economics were set on a price that no longer exists. A creator tool priced around a five-cent second of video cannot simply pass through a doubling of its input cost without changing what it charges or what it delivers.

The tell is how much of the stack was pinned. Teams that abstracted their video calls behind a small internal interface, with the model name in one config file, treat October 22 as a one-line change. Teams that scattered veo-3.1-lite-generate-preview across a dozen services treat it as an archaeology project. The abstraction that felt like over-engineering when it was written turns out to be the thing that decides how the deadline feels.

The platform is not the model

The most useful lesson here is buried in a small detail. The deprecation table pairs each Veo preview with the Omni replacement and says nothing about where Veo goes next. But Veo is not dead. It continues on the Gemini Enterprise Agent Platform, where versioned Veo endpoints carry retirement dates of November 17, 2026 or later, with no replacement named.

In other words, the same model family is being retired on one Google surface and maintained on another. What is being deprecated is not the model. It is the offering: video generation on a preview-gated developer API, priced per second, sold to whoever wanders in. Google is moving video upmarket to the enterprise platform and closing the casual door behind it.

For developers, that reframes the decision. The question is no longer "which video model is best." It is "which vendor's roadmap am I willing to be a tenant of." The model you build on today may be fine. The surface it lives on is the part that disappears.

What to actually do

Three practical moves, in order. First, inventory every Veo call in your stack and note which variant each one uses, because the migration cost is very different for standard, Fast, and Lite. Second, prototype against the Omni replacement before the deadline, not after, because a per-token bill behaves differently from a per-second bill at low and high volumes alike. Third, and least urgent but most important, write down what you would do if this happened again next year. The October 22 deadline is not unusual. It is the new normal, and the teams that survive it are the ones that treat model choice as a lease with an expiry date rather than a purchase.