Gemini Omni 1.2 Flash Generates Video Up to 15 Seconds and Is 30% Cheaper

The biggest pain point of working with video models is familiar to anyone who has made ads or demo videos: you generate a clip, notice that the hero's face has drifted or the logo on the packaging has turned into mush, and you have to pay for a new generation from scratch. Google is trying to solve exactly this problem in Gemini Omni 1.2 Flash.

What's New

The main change is length. The model produces clips of up to 15 seconds natively, rather than stitching them together from pieces, as was the case with the 10 seconds in Omni 1.1. Duration is set in one-second steps from 3 to 15. For social ads and short explainers, this means an entire story fits into a single generation.

The second change is price. Google cut it by 30%: $0.07 per second of 720p video. A 15-second clip will cost a little over a dollar. By comparison, the same volume used to cost noticeably more, and you had to pay again after every failure.

How It Works in Practice

The model accepts references: images of a character, product or style, short motion clips of up to three seconds, and audio cues. Say you have a photo of a model and shots of three outfits from the front and back. From these you can build a 15-second virtual try-on clip with a 360-degree turn, and the cut of each garment should be preserved. You can likewise turn existing catalog photos into spinning product cards or branded ads.

The second practical feature is multi-turn editing. You take a finished clip and say: change the background to a forest, then to a castle, change the lighting or the aspect ratio. Characters and camera angle stay in place. For marketers, this is a way to quickly make clip variants for different audiences without a repeat brief.

Google describes the economical workflow like this: drafts in 360p or 720p, targeted detail edits, then upscaling the successful versions to 1080p or 4K. In other words, the expensive resolution is only used at the final step.

Speech, Camera and Availability

Google claims more natural dialogue, better lip sync, and assigning voices to multiple characters from references. The camera has also become more controllable: long FPV-style flythroughs and pull-backs without extra cuts. Provisioned Throughput is available for enterprise, and the output is marked with an invisible SynthID watermark.

The model is available in the Gemini API and Google AI Studio, in the Google Cloud console, in Flow, Google Vids and Google Ads. In the Gemini app, 15-second clips are so far rolling out to AI Ultra subscribers.

Limitations

All the examples in the post were chosen by Google itself and there are no independent tests yet, so the claims about zero drift and no stitching should be verified on your own data. The company itself admits that fast sports action, abrupt scene changes and complex physics are not yet fully polished. Motion references are limited to three seconds.

What to Do

If you are building a product around generating ad or catalog videos, it is worth recalculating your unit economics: 30% off plus editing instead of regeneration can noticeably reduce the cost of a successful clip. Start by running your real references — packaging, fonts, faces — at 720p and see how stable the consistency is before building the model into production. Separately, check how the SynthID watermark fits your content labeling requirements.