Making video with Veo, in Flow or through the API

Veo is Google DeepMind's video model. You describe a shot in words, or hand it a still to start from, and it returns a short clip with the sound already in it, such as footsteps, rain or a line of dialogue if the prompt asks for one. The current version, Veo 3.1, makes clips of 4, 6 or 8 seconds, wide at 16:9 or upright at 9:16, at 720p unless you ask for more. Higher resolutions need the full 8 seconds, with 1080p on every tier and 4K on the two upper ones. Eight seconds is short. A film made with Veo is therefore a set of separate shots cut together afterwards, the way a live-action film is, and nearly everything in this module is about getting each shot right before it reaches the edit.

Google sells Veo through four doors. Each one bills differently. Flow, at flow.google.com (the old labs.google/flow address now redirects there), is a web studio paid for in credits that come with a Google AI plan. The Gemini app is a chat window, and Google's own page for it says Gemini Omni is replacing Veo there, so the app is not a dependable way into this module's model. Google AI Studio is the developer playground where you make an API key, and the Gemini API is the metered route underneath it, charged per second of finished video to a prepaid Cloud balance.

Veo 3.1 comes in three tiers. The API names them veo-3.1-lite-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-generate-preview, and charges $0.05, $0.10 and $0.40 a second at 720p with sound included. At 1080p Lite rises to $0.08 and Fast to $0.12 while Standard stays at $0.40, and 4K costs $0.30 on Fast and $0.60 on Standard. Flow calls the same three Lite, Fast and Quality, and charges 10, 20 and 100 credits a clip on a non-Ultra plan, counted per generation, and one request can produce two clips. The cheap tier also does less. By Google's API docs, only Fast and Standard take reference images or extend a clip, so Lite is the drafting tier and the upper two are for work a client will see. The same docs list a first and last frame on all three tiers.

Whichever door you use, a request carries the same few decisions. You choose the prompt, the tier, the length, the shape of the frame and the resolution, and then, if you want one, a picture to work from: a still for the opening frame, up to three reference images of a product or a person, or an earlier clip to extend. Flow sets these with buttons. In code each one is a named field, and the API answers at once with a job ticket rather than a video, because a clip takes between 11 seconds and 6 minutes to render by Google's own figures. Your script checks the ticket until the job is done, then downloads the file. Do that promptly, since Google keeps generated videos for two days and then deletes them.

Here is how a small studio would use the two doors that matter, Flow and the API, on one job, a 30-second product film cut from four 8-second shots. They find the look in Flow first. Twelve Fast clips of the opening shot cost 240 credits, about a quarter of a Pro plan's monthly allowance, and judging a prompt by eye is quicker in a browser than in a terminal. Once the art director has picked a direction, the winning prompts move into a script. Drafts run on Lite at 720p and 4 seconds for 20 cents each, so twenty variations cost $4. The four chosen shots are then made at 1080p and 8 seconds on Fast, at $0.96 a take, and two takes of each come to $7.68. Standard at the same settings costs $3.20 a take, which is worth paying for the one shot the whole film depends on.

One more change arrived in 2026. Google's video overview for developers now tells you to use Gemini Omni Flash (gemini-omni-1.1-flash) as the default and to reach for Veo 3.1 when you need scene extension, last-frame control or an older pipeline. Omni Flash costs roughly $0.10 a second at 720p by Google's own estimate, and Flow offers it in clips of up to 10 seconds. Last-frame control is the odd one out. When we called both on 25 September, the API refused Veo's last frame on every tier and Omni's worked first time. We lead with Veo because most of this module's tests ran on it, and Omni gets a lesson of its own later in the module.