Google's Gemini Omni 1.2 Flash adds 15-second clips and multi-turn video editing
Google DeepMind
- Models
- Industry
- Developer tools
- #gemini
- #video generation
- #google deepmind
- #synthid
Google DeepMind announced Gemini Omni 1.2 Flash, a video generation model with native clips up to 15 seconds, multi-turn conversational editing, stronger reference adherence, and a stated 30% price cut to $0.07 per second for 720p. It is described as generally available on the Gemini API and Google Cloud, with reach into Flow, Google Vids, Gemini app and Google Ads.
What Google announced
Google DeepMind introduced Gemini Omni 1.2 Flash, a video generation model it describes as generally available on the Gemini API and Google Cloud, with rollout across Flow, Google Vids, the Gemini app and Google Ads. Three changes stand out. Native clips now run up to 15 seconds, up from 10 in Omni 1.1, and duration can be set in one-second increments from 3 to 15. Clips can be edited across multiple conversational turns, so a user can swap a product, background or aspect ratio without regenerating from scratch. And Google says reference adherence improved: character, product or style images, motion clips up to three seconds, and audio cues are held more consistently across angles and transitions. Google also states a 30% price reduction to $0.07 per second for 720p, and that it drafts at 360p or 720p before upscaling to 1080p or 4K. Outputs carry a SynthID watermark.
Where a marketing team could try it
A short-form video team producing catalog ads could test whether multi-turn editing actually reduces rework. The premise is Google's claim that a detail can be corrected with a follow-up edit instead of a full regeneration, and that product references stay consistent. A plausible trial: take an existing product photo shoot, generate a fixed set of 10- to 15-second clips with a locked first and last keyframe, then run a scripted series of edits (background swap, aspect ratio change) on the same clip. A content producer would judge whether the edits preserve the product's packaging typography and colors well enough to be used without a manual pass in an editor.
How to run a fair trial
A team could start by fixing a baseline: the same briefs run on Omni 1.1 or the current production process, scored by reviewers on reference fidelity, drift between first and last frame, and edit count needed to reach an acceptable cut. Then measure cost per accepted clip and time to acceptance, not just generation time. Include failure checks, such as deliberately complex physics or fast motion that Google itself lists as still under development, and track how often those clips are rejected. A human reviewer should sign off on any clip before it reaches customers, particularly where brand assets or likenesses appear, and teams should confirm current pricing, quotas and terms in the official documentation rather than relying on announcement figures.
Trade-offs and when to hold back
The efficiency argument depends on multi-turn edits genuinely replacing regeneration, since every extra generation carries cost even at a lower per-second rate. Fifteen seconds is still short for narrative work, so teams needing longer sequences must stitch clips and check continuity across them. Reference adherence helps consistency but does not guarantee factual or brand accuracy, and SynthID marking does not remove the need for disclosure or licensing review. Teams should check regional availability, rate limits and enterprise terms before committing a pipeline. If the work relies on fast sports action, rapid script changes or complex physics, Google's own note that development continues suggests waiting or keeping a fallback path.
Context and limitations
- Based on a short excerpt of Google's announcement; the full article and linked documentation were not analyzed.
- Performance, consistency, pricing and availability claims are Google's own and were not independently tested.
- Google notes development is ongoing for fast sports action, rapid script changes and complex physics.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.