Google launches Gemini Omni 1.2 Flash as a generally available video model that creates and edits clips up to 15 seconds. The release adds conversational revisions and lowers the price of 720p generation by 30% to $0.07 per second.

 

The November 5 rollout puts the model in the Gemini API, Gemini for business and Google Cloud production services. Google is also integrating it into Google Vids, Flow, the Gemini app and other creative products.

 

The release changes four parts of Google's video workflow:

  • Clip length rises from 10 to 15 seconds.
  • Users can revise the same clip through follow-up prompts.
  • Image, motion and audio references guide generation.
  • Drafts can be upscaled to 1080p or 4K.

 

Google Launches Gemini Omni 1.2 Flash for Production

Google describes Omni 1.2 Flash as a production-ready update rather than an experimental preview. Developers can call it through the Gemini API or Vertex AI, while companies can use it through Gemini for business and Google Vids.

 

The model accepts text, images, audio and video as references and produces video with sound. Google says it can hold characters, products, clothing and visual styles more consistently across changing camera angles and motion.

 

Those quality claims come from Google's own demonstrations. The public announcement does not provide an independent benchmark comparing reference adherence, motion stability or audio quality with rival systems, so customers will need to test their own workloads.

 

Fifteen-Second Clips and Conversational Editing

Gemini Omni 1.2 Flash lets users set durations in one-second increments from three to 15 seconds. The longer ceiling gives product explainers, social advertisements and short narrative scenes more room than the 10-second limit in Omni 1.1.

 

Multi-turn editing is the more consequential workflow change. A user can ask the model to replace a background, alter lighting, change the weather, swap a product or switch the aspect ratio while continuing from the existing generated clip.

 

Google says the system can combine character or product images with motion clips of up to three seconds and audio cues. Audio references remain restricted to allowlisted customers, which limits one advertised input mode at launch.

 

More on This Story

 

Lower Pricing Changes Video Production Economics

At $0.07 per second, a full 15-second 720p generation has a base model cost of $1.05 before retries, edits or other platform charges. The 30% reduction matters because generative-video projects often require several drafts before a usable result appears.

 

Google supports draft generation at 360p and 720p, followed by upscaling selected clips to 1080p or 4K. That process lets creators spend less while exploring an idea and reserve higher-resolution processing for approved versions.

 

Conversational editing could also reduce waste if a targeted correction preserves the rest of a clip. The practical savings will depend on whether revised videos keep subjects, movement and audio intact without introducing new errors that force another complete generation.

 

The model is aimed at high-volume commercial work, including virtual try-ons, product rotations and advertising variants. Provisioned Throughput is available to enterprise customers that need predictable capacity rather than shared on-demand access.

 

The Gemini Omni Model Card Leaves Evaluation Gaps

Google DeepMind published a model card alongside the release. It confirms the model's inputs, output channels and distribution, but refers readers to earlier documentation for training data, known limitations, acceptable-use rules and detailed safety evaluation.

 

For frontier-risk assessment, Google relies on evaluations of Gemini 3.1 Pro and Gemini 3.7 Flash. The company argues that Omni 1.2 Flash adds no material capability increase that would make it likely to cross a tracked or critical capability threshold.

 

That inherited assessment addresses frontier risk, but it does not substitute for release-specific evidence about visual errors, identity consistency or misleading synthetic media. Google says production filters are applied and every output carries an imperceptible SynthID watermark.

 

General availability gives developers a chance to measure those tradeoffs in real applications. The most useful evidence will be repeatable tests showing how often targeted edits preserve the original scene, how reference fidelity changes over 15 seconds and how total project costs compare after retries.

 

Feature image: Illustrative television-production camera photographed by Panelxf, cropped and resized. The original was released under the CC0 1.0 public-domain dedication.