Gemini Omni 1.2 Flash: 15-second video and 30% cheaper

Gemini Omni 1.2 Flash: 15-second video and 30% cheaper

On Thursday (November 5), Google released Gemini Omni 1.2 Flash, a new version of its video model that generates and edits clips with audio. The maximum clip length went from 10 to 15 seconds and the price dropped 30%, to $0.07 per second at 720p.

For anyone making short video, those 5 extra seconds matter more than they look: 15 seconds is the length of a whole product ad, hook, demo and call to action, in a single generation. I ran the same prompt on 1.2 and 1.1 to see whether the promise holds up.

What changed from Omni 1.1 to 1.2?

Length, price and fidelity to the reference. According to Google's announcement and the Google Cloud post, the model now:

• generates 3 to 15 seconds, chosen second by second (1.1 stopped at 10);
• costs 30% less than 1.1;
• holds face, clothing, packaging and brand color better across camera angles when you send a reference image;
• supports multi-round editing: swap the product, the background, the lighting or the aspect ratio without regenerating everything;
• has more natural speech, cleaner lip sync and distinct voices when more than one person is talking;
• does continuous camera moves (first-person flight, orbit, pull-back) with fewer unwanted cuts.

Google itself admits where it still falls short: fast sports action, quick script changes and complex physics. Every video comes out with an invisible SynthID watermark.

What it really costs

$0.07 per second at 720p, or about $1.04 for a 15-second clip. The Google AI Studio pricing page spells out the rest:

Resolution	Price per second (API)	15 s clip
360p (draft)	about $0.023	about $0.35
720p	about $0.07	about $1.04
1080p (upscaled)	about $0.104	about $1.56
4K (upscaled)	about $0.209	about $3.14

The row that matters is the first one. Google recommends drafting at 360p or 720p, fixing with edits instead of regenerating, and only upscaling at the end. In practice, three 360p drafts cost the same as one 720p clip. Here is how I would work: test the hook and the framing cheaply, and only pay for resolution on what is already approved.

Real test: the same prompt on 1.2 and 1.1

I sent this prompt, in English, to both versions, at 9:16 and 720p, through the Cut.Pro Generator: a woman in a small apartment kitchen, morning window light, holds a bag of coffee up to the camera, says in Portuguese "Esse café aqui mudou a minha manhã" ("This coffee changed my morning"), pours coffee from a moka pot into a white mug and takes a sip. Single take, no cuts. It is the classic testimonial-style product video.

Omni 1.2 Flash, 15 seconds. It cost 255 credits and took about 1 minute and 20 seconds from request to finished file.

Omni 1.1 Flash, 10 seconds (its maximum). It cost 230 credits and took about 50 seconds.

What I saw:

• Price per second. In the Generator, 1.2 came to 17 credits per second and 1.1 to 23. That is 26% less, close to the 30% Google announces. For 25 more credits, 1.2 delivered 5 more seconds of video.
• Staged scene. In the first frame of 1.2, the moka pot and the white mug are already on the counter, ready for the action the prompt asks for later. The framing is an outstretched-arm selfie, which fits "handheld camera". In 1.1, the moka pot shows up in the background and the bag fills half the screen, closer to the lens.
• Packaging text. Both invent the label. 1.2 wrote "COFFEA" with tiny illegible letters underneath; 1.1 wrote an equally meaningless "MUTINA". If the product is yours, send a photo of the packaging as a reference. Generating a label from the prompt still does not work.
• The AI look. 1.2 has polished skin and hair, with commercial-style golden light. 1.1 looks more homemade, with a plainer kitchen. For a testimonial that should look like the customer filmed it, 1.2 needs a prompt that asks for imperfection.

Both videos are up above, from the same request, so you can judge the speech and lip sync by your own standard.

What changes for people who post short video

The 15 seconds and the lower price make it viable to generate variations, not just the final video. Three uses where I see a direct gain:

1. Product ad in one generation. Before, 10 seconds forced you to join two clips and hope the character did not change faces at the cut. Now the hook, the use and the call to action fit in a single clip.
2. B-roll for clips. If you cut podcasts or livestreams, you can generate an illustration of what is being said (the product, the place, the situation) without stock footage.
3. Recurring persona. Character reference accepts up to 6 images, with up to 3 per character for up to 2 characters, according to the API documentation. That is what keeps the same person across different videos, the same principle as the AI persona.

The catch: 100% generated content still competes for attention with real people filming, and platforms are giving more and more weight to recorded material, as we showed in provenance as a signal. AI video works better as support footage or an ad than as a replacement for your face.

Limits the documentation shows

The documentation has details the announcement does not highlight:

• First and last frame only work on clips of 3 to 10 seconds. Asking for 15 seconds with both frames returns an error.
• Extending a clip runs 3 to 10 seconds per round, up to 30 seconds in total.
• Reference audio is not yet available through the API; on Google Cloud, it depends on access approval.
• Recognizable people: images of some well-known people are not accepted as uploads or for editing.

Where you can use it today, and whether it works in Brazil

The model is available on the Gemini API, Google Cloud, Flow and Google Vids (15-second clips at no extra cost for anyone with a Google or Workspace account). In the Gemini app, the rollout starts with Google AI Ultra subscribers. The announcement does not list countries, and the only regional restrictions in the documentation are for the European Economic Area, Switzerland and the United Kingdom. For Brazil, there is no declared block.

It is also already in the Cut.Pro Generator, alongside other video, image and voice models. There you can generate the clip and take it straight to the editor, add captions and schedule the post on social networks, without downloading and uploading files. It is the argument I made in the workflow became the product: the model changes every few months, the path to the post stays.

If you use 1.1 through the API, do you need to do anything?

Check the output, not the code. Starting December 3, 2026, calls to gemini-omni-1.1-flash will be served by 1.2 automatically and billed at the new price, according to the documentation. The bill gets cheaper, but the video changes: a prompt tuned for 1.1 may come out different. It is worth running your five most common generations on 1.2 before the switch, the same lesson that came out of the Sora API shutdown.

What to do starting today

If you already generate product video or support footage, swap 1.1 for 1.2 now: it is cheaper per second and fits the whole ad. Draft cheap, send a real photo of the product as a reference instead of describing the label, and ask for imperfection in the prompt when the video needs to look like real people filmed it. If you do not use generated video, nothing changes in your clipping routine: the gain here is for people producing new pieces, not for people cutting what has already been recorded.