Veo 2

What it is

Veo 2 is a video-generation model that Google announced on 16 December 2024.^[1] It belongs to the family of generative models that turn a text prompt into a moving image sequence, and Google presented it as a step forward in how well such a model handles the physical world and the way people move and look.^[1]

What Google said it improved

When announcing the model, Google cited improved understanding of real-world physics and the nuances of human movement and expression.^[1] Those two things are related but not the same. Physics is about whether objects behave plausibly: whether a glass that is knocked over falls the way glass falls, whether poured liquid behaves like liquid, whether a bouncing ball loses height in a believable way. Human movement and expression are about the small details of faces and bodies — how an expression shifts while someone speaks, how weight moves when a person walks. Both are hard for generative models because a video has to stay consistent from one frame to the next, and errors that would pass unnoticed in a single still image become obvious once the image moves.

Resolution, length and evaluation

Google said Veo 2's outputs reach resolutions up to 4K and can be extended to minutes in length.^[1] The company also said human raters judged it state of the art against leading video models.^[1] That kind of comparison rests on people watching clips side by side rather than on a single number, which is common in video generation because there is no agreed automatic measure of whether a clip looks right.

Where people could try it

At announcement, Veo 2 was added to Google Labs' VideoFX, with expanded access through a waitlist.^[1] In April 2025 it came to the Gemini app for advanced users^[2], and Veo 2 video generation was extended to Gemini app users subscribed to Advanced.^[4] What those users got was narrower than the announced capability: clips in the Gemini app are 8 seconds long at 720p in 16:9 landscape format, and users face a monthly limit on how many clips they can generate.^[4]

The API version

For developers, the model appeared in the Gemini API as veo-2.0-generate-001, released on 9 April 2025.^[3] That endpoint has a shutdown date of 30 June 2026.^[3] Google's deprecation page lists gemini-omni-1.2-flash, or the GA models on the Gemini Enterprise Agent Platform, as the replacement.^[3] The same page says the stated shutdown date is the earliest possible date, with formal notice to follow^[3] — so the date marks the beginning of a window rather than a fixed end point that cannot move.

What its run shows

The short public history of Veo 2 shows how differently one model can be packaged. The announcement described 4K output and clips running to minutes^[1]; the consumer service offered 8-second, 720p, landscape-only clips under a monthly cap.^[4] The two descriptions do not contradict each other — they describe different products built around the same model — but read together they make a useful distinction: what a model can produce and what a shipped product hands to a user are separate questions. The API timeline points the same way, with a named version, a published release date, a published shutdown date and a named successor.^[3] For anyone keeping track of video generation, Veo 2 is a marker of where the field stood at the end of 2024 and of how quickly the surrounding products and endpoints were rearranged afterwards.