Gemini Omni 1.2 Flash: Google's Video Model Gets Cheaper and Longer
Gemini Omni 1.2 Flash is now generally available, generating up to 15-second clips with synced dialogue and sound at 7 cents per second.
What is Gemini Omni 1.2 Flash?
Gemini Omni 1.2 Flash is Google’s updated video generation model, now generally available through Google AI Studio, Vertex AI, and the Gemini API. It generates clips up to 15 seconds long (up from 10 seconds in the prior version), at any length you choose between 3 and 15 seconds, and it does so with synced dialogue, sound effects, and multi-speaker voice matching built into the output. Pricing comes in at 7 cents per second of 720p video, which Google describes as roughly 30% cheaper than before.
TL;DR
- Clip length now tops out at 15 seconds, up from 10 in the previous Gemini video model, and you can set any exact duration between 3 and 15 seconds.
- Pricing lands at 7 cents per second for 720p output, which Google frames as a 30% cost reduction versus the prior version.
- Resolution workflow lets you draft cheaply at low resolution first, then upscale winning takes to 1080p or 4K instead of paying full price on every attempt.
- Multi-speaker dialogue is a standout feature: the model can take a single scene image plus short voice references for multiple characters and produce lip-synced, distinct voices for each one across a 15-second clip.
- Camera and shot consistency improved enough to produce single continuous takes, like a drone flight with zero jump cuts, something earlier video models struggled to hold together.
- Access points include Google AI Studio for quick testing, Vertex AI for enterprise use, and the Gemini API for developers who want to script generation with Google’s SDK.
- Editing through conversation is supported, meaning you can iterate on a generated clip with follow-up prompts (swap a background, change a setting) rather than starting over from scratch.
What changed from the previous version?
The jump from Gemini’s earlier video model to Omni 1.2 Flash isn’t a total overhaul so much as a tightening of the things that made AI video generation frustrating. Clip length increased from 10 to 15 seconds, and instead of being locked into fixed durations, users can now dial in any length between 3 and 15 seconds depending on what the shot needs. Google also says the model holds up better across comparisons with its own 1.1 predecessor: in a side-by-side test using an identical prompt and reference image, 1.2 matched source material (like a poster’s layout and text) more closely than 1.1 did.
The pricing drop to 7 cents per second for 720p output matters here too. Video generation has historically been one of the more expensive AI modalities to experiment with, since a single unsatisfying clip can burn through a chunk of budget before you get something usable. A 30% price cut, combined with the ability to draft at low resolution and only upscale the clips worth keeping, changes the economics of iterating on video prompts.
How does the dialogue and sound syncing work?
One of the more notable capabilities demonstrated with Omni 1.2 Flash is multi-speaker dialogue with lip sync. In one example, Google fed the model a single scene image (a glass conservatory full of orchids and ferns) along with character mouth references and short voice samples for four separate characters. The model then generated a 15-second animated scene where each character spoke in its own matched voice, with lip movement synced to the dialogue, as the characters bickered over who was “copying” whose color palette.
This is a meaningfully different task than just generating a visual sequence. The model has to track which character is speaking at which moment, apply the right voice to the right mouth, and keep the timing believable across camera cuts within the same 15-second window. Testing this directly in Google AI Studio with a custom prompt describing four characters and their lines produced a comparable result: a short scene with distinct voices attached to the correct characters and lip sync that held up across the clip.
Can it hold a single continuous shot without cuts?
Yes, and this is one of the more visually impressive capabilities shown. In one test, a single text prompt asked for a continuous 15-second drone shot flying past a mountain summit, down a glacier valley, and into an alpine meadow, with no cuts. The model produced exactly that: one unbroken camera move rather than a series of stitched clips.
● Reading src/server.ts
● Editing 3 files
+ import { limit } from 'rate'
- app.use(cors())
✓ Tests passed · ready for PR
A second example pushed this further with cinematic camera control guided by architectural reference photos. The model generated a continuous first-person drone path flying across a piazza, through bronze doors, and up into a dome, again with zero jump cuts. Maintaining spatial consistency and smooth motion across a sustained single take is historically one of the harder problems in AI video generation, since models tend to lose track of geometry or introduce visual artifacts the longer a shot runs. Omni 1.2 Flash’s ability to hold a clean continuous shot for the full 15 seconds suggests meaningful improvement in temporal and spatial consistency.
How does identity and object consistency hold up?
Another demonstrated use case was a virtual outfit try-on: Google combined a single model photo with front and back reference shots of three different outfits, and the model generated a 15-second sequence showing 360-degree turns and match cuts between outfits. The claim is that the model keeps the person’s identity and each garment’s details consistent throughout, and in the demonstrated example this held up reasonably well, with fabric texture and fit staying stable across camera angles.
A related example showed a poster animation built “layer by layer” from a single reference image, with Google directly comparing 1.1 and 1.2 output on the same prompt. The newer version reportedly matched the reference image’s text and materials more closely, suggesting the model is better at respecting visual details from source images rather than drifting from them over the course of a generation.
Is Gemini Omni 1.2 Flash worth using for video projects?
For anyone prototyping short-form video content, ads, social clips, or scene concepts, the combination of lower per-second pricing and a low-resolution draft workflow makes experimentation more practical than it was before. You can test a prompt cheaply, confirm the camera move or dialogue timing works, and only pay for a 1080p or 4K upscale once you’re satisfied with the result.
That said, this is still a 15-second-clip model, not a tool for generating long-form video in one pass. Multi-turn editing (prompting the model to swap a background or change a setting within an existing clip) helps extend what you can do without regenerating from scratch, but anyone expecting minute-long continuous narratives will need to stitch multiple generations together manually. For short, high-concept clips, especially ones involving dialogue, voice matching, or a single uninterrupted camera move, the model performs at a level that looks genuinely competitive with hand-produced short video.
Frequently Asked Questions
What is the maximum clip length for Gemini Omni 1.2 Flash?
Clips can run up to 15 seconds, and you can specify any exact length between 3 and 15 seconds rather than being locked into fixed presets.
How much does Gemini Omni 1.2 Flash cost?
Google prices it at 7 cents per second of generated video at 720p resolution, which the company describes as about 30% cheaper than the previous version.
Where can I use Gemini Omni 1.2 Flash?
It’s available through Google AI Studio for quick testing, Vertex AI in Google Cloud for enterprise workflows, and the Gemini API for developers who want to script generation using Google’s SDK.
Can Gemini Omni 1.2 Flash generate video with dialogue and sound?
Yes. It can take a scene image plus short voice references for multiple characters and generate a clip with synced, distinct dialogue and lip movement matched to each character, along with ambient sound and music.
Does it support higher resolutions than 720p?
Nothing gets lost.
Yes. The recommended workflow is to draft at lower resolution first, then upscale the final clip to 1080p or 4K once you’re satisfied with the result, which keeps iteration costs down.