Google releases Veo 2

Background

A text-to-video model takes a written description and turns it into a moving clip. It is a harder problem than generating a still image: the model has to keep the same objects and people consistent from frame to frame, and it has to produce motion that looks physically believable rather than melting or sliding. Google has published its generative media models through Google Labs, wrapping them in small experimental apps — VideoFX for video, ImageFX for images — that users reach through a waitlist rather than an open product.^[1]

What Google announced

On 16 December 2024, Google announced Veo 2. The company said the model could produce video at up to 4K resolution, with improved handling of physics and of human movement.^[1]

The release was not a single model but a package. Veo 2 rolled out in VideoFX, the video experiment, alongside an improved Imagen 3 in ImageFX, the image experiment; Google said the new Imagen 3 in ImageFX was rolling out to more than 100 countries.^[1] Alongside both, Google introduced Whisk, a tool from Google Labs that combines Imagen 3 with Gemini, and said it was launching in the United States the same day.^[1] In other words, the video model, the image model and a Gemini-backed tool for working with images all arrived together.

Getting access

Access at launch was not open. Visitors were asked to sign up for the VideoFX waitlist through Google Labs, so the first users of Veo 2 got in through a queue rather than through a general release.^[1]

The move into a mainstream product came later. Veo 2 reached the Gemini app in April 2025,^[2] and in that month video generation was extended to Gemini app users subscribed to Advanced.^[3] The version inside the Gemini app was much more modest than the announcement numbers: clips generated there are 8 seconds long at 720p.^[3]

Why it matters

The episode shows how generative video was being delivered at the end of 2024 and into 2025 — not as a standalone product but as one component of a media stack. A video model, an image model and an assistant-based tool (Whisk, which combines Imagen 3 with Gemini) were announced in the same breath, and all of them sat inside Google Labs experiments that users entered through a waitlist.^[1]

It also shows the gap between what a model can do in a demonstration and what a product ships. At announcement, the headline claim was video at up to 4K resolution.^[1] By the time Veo 2 arrived inside the Gemini app, the output was 8-second clips at 720p, and only for Advanced subscribers.^[3] The high-resolution capability and the consumer feature were not the same thing: the product form was short and gated behind a subscription, while the research claim was about resolution and physical plausibility.

For users, the practical change was that video generation became a feature of a general assistant rather than a separate experiment — something reached from the same app used for chat, available to paying subscribers. That shift from a Labs waitlist to a subscription tier is the part of the announcement that carried through into how people actually used the model.^[2][3]