There’s a new model to choose from in the ChatX video generator: Gemini Omni 1.2 Flash from Google. It creates clips of up to 15 seconds in one go, always with sound. Google made the model generally available on November 5, 2026. Here’s what it can do, how many tokens a clip costs, and how to write your prompts so the result turns out the way you want.
The key facts at a glance
- Length: 3 to 15 seconds per clip, in one-second steps. Its predecessor, Omni 1.1, topped out at 10 seconds.
- Sound: Speech, sound effects, and music are generated together with the picture. There is no such thing as a video without an audio track.
- Resolution: 360p, 720p, 1080p, and 4K. The two highest settings are upscaled.
- Format: Landscape 16:9 or portrait 9:16, 24 frames per second, as MP4.
- Input: a text, a starting image, or a starting and an end image. With two images, the maximum is 10 seconds.
- Reference images: Up to six images keep a character or a product consistent throughout the clip.
- Labeling: Every video carries the invisible SynthID watermark, which makes AI videos identifiable.
According to Google’s announcement, dialogue sounds more natural and lip movements match the speech better than before. The model is also 30 percent cheaper than Omni 1.1.
From Veo to Omni
If you’ve worked with Google’s video models before, you know the name Veo. In the Gemini API, Omni has taken over that role:
- June 30, 2026: Omni Flash launches as a preview, with clips of up to 10 seconds in 720p.
- August 27, 2026: Omni 1.1 Flash becomes generally available and adds starting and end images as well as resolutions up to 4K.
- October 22, 2026: Google shuts down Veo 3.1 in the Gemini API and names Omni as its successor.
- November 5, 2026: Omni 1.2 Flash follows with 15 seconds, up to six reference images, and a lower price.
What a clip costs on ChatX
You’ll find Gemini Omni 1.2 Flash in the video generator alongside Seedance 2.5, Kling 4.0, Wan 3.0, MiniMax H3 Max, and HeyGen Video 1. You need a free account to use it. Billing is per second of video, and the price depends on the resolution.
| Resolution | Tokens per second | 10-second clip |
|---|---|---|
| 360p | about 5,700 | about 57,000 |
| 720p | 17,000 | 170,000 |
| 1080p | 25,500 | 255,000 |
| 4K | 51,000 | 510,000 |
So a clip with the full 15 seconds in 720p costs 255,000 tokens. Our advice: try out an idea in 360p first. That costs a third as much and shows you whether the composition, motion, and sound are right. The higher resolution is only worth it once the prompt is dialed in.
Our test with three short clips
We tried three types of input: text, starting image, and reference image, each with 3 seconds in 360p. That’s the cheapest setting and costs about 17,000 tokens per clip.
| Input | Request | Wait time |
|---|---|---|
| text only | red paper kite over a meadow | 24 seconds |
| starting image | camera moves toward the storefront window of a bookstore, the shop door opens | 28 seconds |
| reference image | the same bookstore on a rainy evening, people with umbrellas walk past | 27 seconds |
All three clips came with an audio track. With the starting image, the video begins exactly at the uploaded photo: the camera moves in, the shop door opens, and the chalkboard with the German text “Lesung heute um 19 Uhr, Eintritt frei” (Reading today at 7 p.m., free admission) stays legible letter for letter.
Clip with a starting image: 3 seconds, 360p, with sound. The bookstore image comes from our Nano Banana 2.1 test.
With the reference image, the model rebuilds the scene from scratch. The shop, the lettering, and the chalkboard were preserved, and rain, wet pavement, and passersby with umbrellas were added. Small details got lost, such as the bench in front of the door. A clear description matters: “It’s raining heavily, wet pavement” works better than just “in the rain”.
Clip with a reference image: the same shop, now in the rain.
How to write good prompts
The following tips come from Google’s prompt guide for Omni Flash. It’s best to write your prompts in English (more on that in the last section).
- Ask for a single shot: Left to its own devices, the model likes to cut several shots together. If you want one continuous scene, add “In a single continuous shot” or “No scene cuts”.
- Describe the sound: Otherwise Omni picks a fitting audio track on its own. You have to ask for music explicitly, for example “Include calm background music”. If nobody should speak, “No dialogue” helps.
- Specify timing: You can define when something happens. That works in plain language (“After 3 seconds …”) or with timestamps such as [0-5s].
- Rule out unwanted things directly: The model has no separate field for negative prompts. Just write what you don’t want into the prompt.
- Specify text in the picture: Signs or on-screen text are most likely to work if you give the exact wording in quotation marks.
Three examples to adapt:
Calm single scene with sound effects
In a single continuous shot, a ceramic mug of coffee steams on a wooden table next to a rain-streaked window. The camera slowly pushes in. Sound design: soft rain, a distant tram bell. No dialogue, no music.
Sequence with timestamps (15 seconds)
[0-5s] A cyclist rides along a foggy lakeside path at dawn. [5-10s] She stops and takes a thermos from her bag. [10-15s] Wide shot: the sun breaks through the fog over the lake. Calm acoustic guitar music throughout.
Product clip with legible lettering (portrait)
A slow rotating shot of a matte green water bottle on a white pedestal, soft studio light. A small sign next to it says: "Refill. Repeat." Upbeat, minimal electronic music. No scene cuts, no people.
For the product clip, you can upload a photo of your own as a reference image. The bottle in the video will then stay much closer to the original.
Where the model hits its limits
- Language: According to Google, only English has been fully tested. Prompts in other languages, such as German, can work, but the results vary more.
- People in reference images: Google rejects photos of specific recognizable people. Images showing minors are likewise not accepted in the European Economic Area, Switzerland, and the United Kingdom.
- Content filters: Google checks the prompt and the finished video. Anything that violates Google’s policies gets blocked. ChatX has no influence on that.
- Difficult subjects: Google itself names fast sports scenes and complex physics as areas that are still being worked on.
- High resolution: 1080p and 4K are produced by upscaling. So they don’t necessarily contain more image detail than 720p.