Google on November 5 released Gemini Omni 1.2 Flash, a video generation and editing model, and made it generally available the same day. The launch blog leads with three changes: the maximum native clip length rises from 10 seconds to 15 seconds, reference images are reproduced more faithfully, and 720p video drops to $0.07 per second, 30% below version 1.1.
The post is credited to Anish Nangia, a product manager at Google DeepMind. Developers can call the model through the Gemini API in Google AI Studio, while enterprise customers use Vertex AI on Google Cloud, which also offers Provisioned Throughput. Consumer entry points are Flow, Google Vids and the Gemini app. In the Gemini app, 15-second video is rolling out gradually and is currently limited to Google AI Ultra subscribers; in Google Vids, 15-second clips generated with a Google account or a Workspace subscription carry no extra charge. Google Ads has also integrated this version to produce video ad creatives in bulk.
Duration and consistency
Clip length can be set anywhere from 3 to 15 seconds in 1-second steps. Version 1.1, released in August, capped a single clip at 10 seconds, so longer videos had to be built by extending a clip segment by segment. Version 1.2 makes the single clip longer while keeping extension and first-and-last-frame interpolation. According to the blog, characters, objects and scenes drift and warp less between the first and final frames.
There are three kinds of reference input: images of characters, products and styles; motion reference clips of up to 3 seconds; and audio prompts. Google's examples are virtual try-on and 360-degree product showcases, and it says the new version keeps faces, clothing, and the lettering and brand colors on packaging consistent across different camera angles.
Multi-turn editing now covers the full 15-second clip. You can swap products, change backgrounds, adjust lighting and weather, or change the aspect ratio, each time building on the previous result instead of regenerating the whole clip. On the audio side, the model adds multi-speaker conditioning, which assigns different voices to the matching characters, and lip sync has been improved. For camera movement, the blog singles out continuous shots such as FPV fly-throughs, pull-backs and orbits, and says there are fewer jump cuts and steadier framing.
All output carries an invisible SynthID watermark.
Doing the math at 720p
Google has published a per-second price for only one tier, 720p; for 1080p and 4K you have to check the pricing page. The workflow the blog recommends is to draft at 360p or 720p, refine through multi-turn editing, and upscale to 1080p or 4K only at the end.
A rough calculation based on the public figures:
| Clip | Version 1.1 (720p, about $0.10/sec) | 1.2 Flash (720p, $0.07/sec) |
|---|---|---|
| 15-second single clip | About $1.50, needs extension | About $1.05 |
| 30-second finished video (two clips) | About $3 | About $2.10 |
The per-second price for version 1.1 is an approximation we back-calculated earlier from the token-based figures in the pricing documentation, while the $0.07 for version 1.2 is a number Google gave directly. The gap between the two roughly matches the 30% cut the blog describes.
The table covers a single generation only. The blog does not say how repeated drafts are billed, whether each round of multi-turn editing is charged for the full clip length, or how much extra upscaling to 4K costs, so real bills will run higher than the table suggests. For teams making e-commerce short videos on a draft-many-times-then-upscale workflow, the cost per video depends mainly on how many rounds of edits they go through; the listed price is only a starting point.
Shortcomings Google lists itself
The blog names three areas still being improved: fast-moving sports action, scripts with frequent dialogue changes and complex physics effects. All performance claims rest on samples Google selected, including a poster sequence and first-and-last-frame comparisons. There are no quantitative benchmark scores and no third-party evaluations.
For developers in mainland China, the Gemini API and Vertex AI are not available, so using them requires an overseas entity and account. Domestic teams already have video models such as Kling and Jimeng, and Kling 4.0 could already generate 30 seconds in a single clip when it launched in October. The pressure from a 15-second clip at $0.07 per second falls mainly on overseas rivals that bill by the second.
Sources: Google official blog, Gemini API pricing page, CocoLoop; figures checked include $0.07 per second at 720p, single clips of 3 to 15 seconds, and the 30% cut versus version 1.1.