Gemini Omni: Features, Pricing and How to Use It

Gemini Omni and Google Veo are not interchangeable labels

Google has used both Gemini Omni and Veo for video generation. In the setup described here, Omni is the newer default model in the Gemini app and Flow, while Veo has been positioned for specific editing tasks like extending scenes or refining a render. Google has also indicated Omni would replace Veo in the Gemini app, and access to Veo through the Gemini API has changed.

The practical takeaway: pick based on what task you're doing and what options appear in your account today, not on a model name from an older tutorial. Model versions and routing can change quickly. A workflow that worked with one model or API version may not match the interface you see now.

Clip length and resolution shape the use case

Generated clips are short - usually around four to ten seconds depending on the workflow. Output options range from 360p previews up through 720p and 1080p, with 4K available on some higher paid tiers. Scene extension can push a clip further in increments, up to a stated maximum of 40 seconds. What you can do depends on the plan and the product configuration.

Those limits matter more than they might sound. A ten-second clip works well as a social post, a visual concept, or one shot within a larger edit. It isn't a substitute for longer narrative sequences or anything that needs a continuous chain of actions. If you need more runtime, you'll likely be assembling or extending scenes - and then reviewing the joins carefully.

Consistency and control still require review

A model can follow your prompt closely and still introduce continuity issues. Between separate clips, characters, objects, or environments can drift. Even small prompt changes can trigger broader visual changes than you expected. That's one of the reasons it pays to iterate and check every render instead of assuming the next edit will preserve everything you liked.

For client work, product marketing, or any clip where accuracy matters, treat the output as a draft until a person reviews it. Verify identity, product details, movement, text, and audio. The real question isn't whether Omni can produce something impressive - it's whether it can match your brief consistently enough to save you time downstream.

Estimate cost by workflow, not by a single render

The first clip you generate is only part of the real cost. In practice, a usable asset may take several attempts: one to find the composition, another to fix movement or lighting, and more if you need multiple versions. Editing can also consume extra usage depending on how the product counts generations and revisions.

Before you scale up, run a small test and track:

• How many generations it takes to land on an acceptable first result.
• Whether editing a clip uses the same allowance as creating one.
• The output resolution and scene-extension options available on your plan.
• How often you need to regenerate or create variations.
• The time required for human review and post-production.

That gives you a more useful cost estimate than comparing monthly plan prices alone. A lower-cost plan may be enough for occasional concepts. If your team is making clips every week, predictable capacity, access controls, and review time usually matter more than the advertised entry price.

A practical first-video workflow

1. Open Google Flow and sign in. Use the Google account associated with the plan or access you want to use.
2. Create a new project. Start with a single, specific scene instead of a whole campaign or multi-shot story.
3. Write the initial prompt. Include subject, setting, main action, and one or two important visual choices. Add reference images if you need the subject, product, or environment to resemble something specific.
4. Review the generation cost before approving. Flow may ask you to confirm that you want to spend credits. Keep approval visible while you're learning the system - automatic approval makes it easy to lose track of repeated experiments.
5. Wait for the render and inspect the whole clip. Don't stop at the first frame. Check for subject changes, awkward motion, unwanted objects, audio problems, and whether the action completes as requested.
6. Make a targeted edit or generate another version. Select the clip and describe one change at a time. If you're trying to alter just one element, ask Flow to keep everything else the same. It helps, but it doesn't guarantee perfect preservation.
7. Download the version you reviewed. Use the download option in the clip controls or media view. If you generate alternatives, track which version has been approved.

Keep the project reviewable

If this is a team workflow, name projects and versions so their status is obvious. A clip that looks finished inside a generation workspace isn't automatically approved for publishing. Record who reviewed it, which version was selected, and whether audio, brand details, and rights have been checked.

That avoids a familiar mistake: someone grabbing the latest file in a shared folder and treating it like the final asset. If your team later connects generation to a content calendar or approval process, keep the distinction between draft, review, and approved output clear.

Start with the shot, then add detail

A good starting prompt covers four basics:

• Subject: who or what is in the shot?
• Setting: where is it, and what's visible in the background?
• Action: what should the subject do during the clip?
• Camera or mood: what is the camera doing, and what visual atmosphere do you want?

For example: "A grey lurcher sprints along a cliffside beach in one continuous tracking shot, with waves breaking in the background." That gives the model a subject, a setting, a clear action, and a camera approach. Generate that base first, then add secondary details like wardrobe, weather, music, or a specific color grade.

Make each revision narrow

If the first render is close, change one thing at a time: "Make the lighting more dramatic" or "Change the clip to an anime style." Broad requests can shift the subject, setting, framing, and motion all at once. Narrow instructions make it easier to tell what improved the clip and what introduced a new problem.

Reference images can help when visual consistency matters - especially for a recognizable product or character. They're not a guarantee, though. The model may interpret details differently, and separate generations can still vary. Use references as constraints to improve the odds, then inspect the output.

For audio, describe delivery instead of relying on typography. If a character should yell, say that they yell, and describe the tone or emotion. Writing dialogue in all caps can confuse audio generation rather than reliably controlling volume.

A prompt checklist before you generate

Before spending credits, ask:

• Is there one main action, or several unrelated events?
• Did I describe the setting and camera movement clearly?
• Which details are essential, and which can wait for an edit?
• Do I have a reference image for details that must stay recognizable?
• Did I specify dialogue delivery in words instead of capitalization?

This isn't about finding a magic prompt formula. It's about making each generation a useful test. A clear base prompt gives you a starting point; a focused revision tells you whether a specific change helped. That feedback loop beats trying to include every possible requirement in one huge prompt.