Gemini Omni 1.2 Flash: AI video with sound and the real cost of an ad
Gemini Omni 1.2 Flash generates up to 15 s of video with sound. API walkthrough, per-second pricing, and what a finished ad costs.
Gemini Omni 1.2 Flash is Google's video-with-sound model that on November 5, 2026 left preview and became generally available in its API. It generates up to 15 seconds of video with audio included, and a 15-second clip at 720p costs about $1.04. Here's the API walkthrough, who it's not for, and what a finished ad costs, not just one attempt.
What changed with Gemini Omni 1.2 Flash
It isn't a new video model: it's the model leaving preview. Google declared it generally available with three changes. It generates up to 15 seconds of video with native audio. It keeps the same character with up to six reference images. And it charges 30% less for output video, with double the usage limits on the paid plan, according to Google's announcement.

Also, version 1.1 gets redirected to 1.2 on December 3, already at the new price, and Veo 3.1 Preview was shut off in the API on October 22. According to Wyzowl, as cited by other sites, 91% of businesses use video for marketing: what changes is what it costs to test an idea.
How it works inside: the identifier
Everything goes through the Interactions API: a single call with text and, if you want, images or video. You set the duration (3 to 15 seconds), the aspect ratio (landscape or vertical), and the resolution. The key piece almost nobody mentions: every video saves an identifier, and with it you request changes without re-uploading. That makes it a conversation instead of a slot machine, but it only works if you leave saving turned on.
The full walkthrough, step by step
- You need a paid account: the free plan can't use this model.
- You request the video with a single call. Example: a jar of artisan jam on a wooden table, morning light, the camera slowly pushing in, quiet kitchen sounds; in the parameters, 10 seconds, vertical, 720p.
- The response carries the encoded video; if it's larger than 4 MB, you ask for it as a file and wait for it to become active before downloading.
- If you don't write "single unbroken shot, no scene cuts", the model may put several shots in one clip. And if you set one duration in the text and another in the parameters, the parameters win.
- To edit, you send the previous video's identifier and say, for example: "change the table to marble, keep everything else the same".

The math: what each clip costs
Google charges video output tokens at $12 per million. At 720p that's about 7 cents per second: 15 seconds cost about $1.04 and 10 seconds, $0.70. At 360p it's about 2 cents per second, so 15 seconds come to about $0.35. Version 1.1 charged about 10 cents per second at 720p, so going from 10 to 7 is 30% less; per million tokens it went from $17.50 to $12, 31% less. What goes up in price is resolution: 1080p costs about $0.104 per second and 4K about $0.209.
Three use cases (our own analysis)
- Online store with six product photos: you pass them as references and ask for a 10-second vertical ad. Four versions to test on Reels or TikTok cost $2.80 in total. The benefit is testing four ideas before paying for advertising.
- Restaurant with a video of its dish: you upload the video through the Files API and ask to swap the background for a terrace at sunset, "keep everything else the same". Editing uploaded videos isn't available in all regions, so check before promising it.
- Before and after: you pass two images and describe the transition; you get a clip of 3 to 10 seconds. Without two good shots with the same lighting, the model invents what happens in between.
How to apply it in your business?
- Test everything at 720p and upscale only the winner.
- Budget four attempts per ad: 4 × $1.04 is $4.16 per finished ad; 20 ads a month, $83.20.
- Set the duration in the parameters, not in the text.
- Write "single unbroken shot, no scene cuts" and describe the sound, because the model adds it by default.
- Leave saving turned on if you plan to edit.
The cost of getting started isn't only money: you need a paid account, an API key, and someone who can run a code call. Google publishes a quickstart notebook. For another Google advance in video with sound, read the full breakdown of Genie 4.
Who it doesn't fit and what you're not being told
It doesn't fit someone who needs their own voice or an exact voiceover (it doesn't accept audio as input or edit voices), someone working with Spanish dialogue (English is the only fully evaluated language), or someone still on the free plan. On the "4K video" claim: on Google's pricing page, 1080p and 4K appear as upscaled, and upscaling doesn't add detail the model didn't generate. Better consistency with six references is Google's claim with no independent proof, and there's no independent measurement of how audiences react. Every video carries an invisible SynthID watermark.
It breaks if you set saving to false (you can no longer edit on later turns), and extension reads only the last 10 seconds. First-tier limits are 8 requests per minute and 40 per day.
Frequently asked questions
What is Gemini Omni 1.2 Flash?
Google's model that generates up to 15 seconds of video with native audio, now generally available in the API.
How much does it cost?
About 7 cents per second at 720p: 15 seconds cost about $1.04.
Is there a free plan?
No: the free plan can't use this model.
Can I edit a video that's already generated?
Yes, with the video's identifier, as long as saving is turned on.
Conclusion
Gemini Omni 1.2 Flash lowers the cost of each attempt at video with sound; what's missing is your idea and your judgment in choosing which of the four works. Do the math on a finished ad, not on one attempt, and test at 720p before upscaling.