Used-car listing videos: Omni drafts $0.30, finals $1.50

For a dealership listing video, animate the lead photo of each car with Gemini Omni Flash 1.1: an 8-second clip is $0.30 at 360p, $1.00 at 720p and $1.50 at 1080p on Sume. Twenty cars at 1080p is $30.00. Run the 360p draft first and keep only the winners for the final render. Prices were read on 2026-10-08.

Do not let the model invent the car. Use the real photo as the first frame, and keep the prompt to camera and light.

The request that works for a car

Send the front three-quarter photo as image_url. Ask for a slow orbit or a slow push-in, and say what must not change: the paint colour, the wheels and the plate. Google's Omni page suggests Keep everything else the same for targeted edits.

• Use 16:9 for a dealership site and 9:16 for a short-video feed; Omni accepts only those two.
• Write the licence plate text out if it must be readable, or avoid showing plates, since model-made text can be wrong.
• If you have a short walk-around clip, send it as a reference video: up to 3 clips of at most 3 seconds each.
• If only the colour is wrong in a finished clip, use the video_url edit mode rather than a fresh roll.

Cost per 8-second clip

The Veo rows are Google's direct list for a 1080p 8-second clip, which Google says needs the full 8 seconds.

Why the Veo rows are not the answer

Veo 3.1 Lite looks cheaper per clip, but Sume does not list any Veo model, so it would be a separate Google account with its own limits and a 2-day file window. Omni is also the model Google's own docs point to for quick image-to-video, and it is the one Sume lists. Check what you need from each: the Omni row on Sume gives 3 to 10 seconds, 10 image references and a video-edit mode in one id.

Whatever the route, say in the listing that the clip is AI-animated from the real photo. Google marks Omni output with SynthID, and some viewers will check.

Accuracy is the rule

A listing video must match the car. Use real photos of the actual vehicle, ask only for camera moves, and never let the model add features, trim or wheels. Review each draft against the photos. If the model changes a color, a plate or a dent, that clip cannot be used, whatever its quality.

Do not generate people into the shot. Google notes certain recognizable people are unsupported, and a seller standing next to the car is better filmed than generated.

Submit with an Idempotency-Key header so a retry after a dropped connection returns the original job instead of a second charge; the same key with a different payload returns 409. Then poll GET /v1/jobs/:id/status and fetch GET /v1/jobs/:id/result when it finishes.