Gemini Omni 1.2 Flash makes longer video edits cheaper, with caveats

Gemini Omni 1.2 Flash extends Google’s video-generation and editing model from 10-second clips to 15 seconds, while adding multi-turn conversational edits and stronger adherence to image, video, and audio references. Google says the Gemini API price for 720p output is $0.07 per second, 30% less than Omni 1.1 Flash. That makes a 15-second 720p generation about $1.05 at the listed rate before any other charges or workflow costs. Google also describes drafting at 360p or 720p, then upscaling selected clips to higher resolutions. The practical change is a longer first pass and cheaper iteration, rather than a demonstrated jump in independently measured video quality.

Google announced the update on November 5, 2026, and describes general availability through the Gemini API and Google Cloud. Its Cloud post lists generation and editing up to 15 seconds, API access, and Google Cloud Provisioned Throughput. Google also says the model is available across a mix of its own products, with different rollout paths: 15-second clips in Google Vids, rollout to Ultra subscribers in the Gemini app, and business workflows through Cloud. These channels matter because a feature listed in the announcement may not be available in every interface or account at the same time. Developers should check the relevant API documentation and access terms before treating the product announcement as a complete implementation specification.

The clearest comparison is with Omni 1.1 Flash: Google states that the previous model supported up to 10 seconds, while 1.2 supports up to 15 seconds and lowers the 720p price by 30%. Google’s launch materials also show examples intended to demonstrate continuity, reference following, and conversational edits, but those are provider demonstrations, not an independent evaluation against competing video models. The new model card identifies Omni 1.2 Flash as based on Gemini Omni Flash and describes video output with audio. It does not provide a separate benchmark table establishing how much better its consistency or quality is than the predecessor. So the verified gains here are concrete product limits and published pricing; the size of the quality improvement remains a provider claim.

There are meaningful constraints for teams deciding whether to build around it. Google says audio-reference input is gated behind allowlisting, and its model card says the model can change people’s speech during editing, a capability Google is currently restricting. Google also flags ongoing work on fast sports action, rapid script changes, and complex physics. Those limitations make it sensible to test the exact footage, reference types, and edits a production pipeline needs, then check cost at the chosen resolution and clip length. SynthID watermarking is described for outputs, but it does not establish that generated content is factually accurate or that a workflow’s rights and review requirements are met. The update lowers the cost of trying a 15-second clip; it does not remove the need to evaluate output consistency and access constraints in the intended use case.