How long can a Gemini Omni Flash video be?

Exact to a rounding of about a hundredth of a second. Asking for 3 seconds yields a file measured at 3.008 s, and the token accounting matches precisely — a 3-second clip consumes exactly three seconds' worth of video tokens. That matters because the price is linear in duration: half the length is half the bill.

It also means a draft pass is genuinely cheap. Testing a shot at 3 seconds costs $0.30; committing to the full 10 seconds costs $1.00. There is no quality penalty for the short version — the same model, same resolution, just fewer frames.

Omni Flash is a world model rather than a video generator with a soundtrack bolted on: it produces the picture and the sound in the same pass, lip-sync included. There is no flag to disable audio, and no price difference for it — it is already in the per-second rate.

You steer the audio the same way you steer everything else: in the prompt. A line like Audio/sound: rain on a metal roof, no music is honoured; there is no separate audio field. If you need a silent clip, generate normally and drop the track in your editor.

Since November 2026 the simplest answer is Omni 1.2: one generation runs up to 15 seconds, set to the second, at a lower price per second than 1.1 — long enough for a complete action or a short line to camera without a second pass. The one exception is a clip that animates between a first and a last frame: Google caps that at 10 seconds on 1.2 too.

Beyond that, extend the scene. Omni 1.1 and 1.2 have an extend task: open a finished clip, choose Extend, describe what happens next, and the model appends 3 to 10 seconds to the end of it. The whole source is read every time — a 13-second result of one extension goes back in as 13 seconds, not trimmed to 10 — so extensions chain, and the chain stops at 40 seconds. Reference images can bring a new subject into the continuation. Extension is billed on the full resulting length, so the last pass of a long chain costs as much as a clip of that length.

The original 1.0 generation, still available as Omni Flash 1.0, rejects the same task — a correctly formed request fails with *This model does not support video extension*. If you need a continuous take longer than 40 seconds, Veo 3.1 extensions add 7 seconds at a time up to 148 seconds, and Wan 3.0 renders 30 seconds in one run.

One more limit worth knowing before you plan around it: video-to-video editing keeps the source length. An edited clip has to come out the same duration as it went in, so editing is not a way to stretch a shot — extension is.