15-Second Gemini Omni Clips and the Real Cost of a 60-Second Ad

What actually changed between 1.1 and 1.2

Google made Gemini Omni 1.2 Flash generally available on 5 November 2026. The launch announcement names three changes that matter to marketers: native 15-second clips, multi-turn conversational editing and better reference following. The 720p price dropped 30% to $0.07 per second, and The Next Web notes it was $0.10 per second in June.

Clip length now runs from 3 to 15 seconds. The 1.1 limit was 10. According to Google Cloud's documentation, the model generates or edits video with synchronized audio at 360p, 720p, 1080p or 4K. Its model ID is gemini-omni-1.2-flash, and the docs list it as accessed through the Interactions API rather than the GenerateContent API.

For context, the Omni line debuted at Google I/O on 19 May 2026. Version 1.1 added scene extension and first-and-last-frame control, and 1.2 adds more control over output resolution, per Crypto Briefing. If you already use the API, you will move to 1.2 whether or not you plan to. The Gemini API changelog, as summarized on DEV Community, says requests to 1.1 route automatically to 1.2 from 3 December 2026 and are billed at 1.2 rates.

• Clip length: 3 to 15 seconds, up from a 10-second maximum
• Editing: multi-turn conversational edits on an existing clip
• Price: $0.07 per second at 720p; video output tokens at $12.00 per million versus $17.50
• Output: 360p to 4K with synchronized audio
• Availability: Gemini API and Google Cloud now, rolling out in Flow, Google Vids, Google Ads and the Gemini app

What a single 15-second clip can carry

Fifteen seconds is a real ad unit, not just a demo length. In Google Cloud's GA post, a partner says 15-second 1080p generations let users create a complete CTV ad in a single generation. Google also positions the model for content businesses make in volume, such as product videos, virtual try-ons and ads.

In practice, one clip holds one idea in one setting. A skincare bottle on a bathroom shelf, a hand picking it up, a close-up of the texture, a final hero shot: that fits. A problem-demo-proof-offer structure with two locations and a voiceover that names a price does not fit well. Once the brief needs a scene change or a before-and-after, you are stitching clips, and stitching brings its own issues.

Some content is still a gamble. Google says it is still improving fast sports action, quick script changes and complex physics, as daily.dev reports. My expectation, given those stated weak spots, is that liquid being poured, a running shoe mid-sprint or a kitchen appliance with visible moving parts will need several takes. Turntable product shots, slow lifestyle scenes and ambient B-roll should be much safer.

Where chat-based editing saves time, and where it doesn't

Most AI video time goes into near-misses. The clip is mostly right, but the light is too cold or the camera moves too fast. With multi-turn editing you can ask for warmer light or a slower push-in on the existing result. Before, you rewrote the prompt and hoped the next roll kept what you liked. For a marketer who is not a prompt specialist, this is the most useful change in the release.

Extensions help too. OrcaRouter's comparison says 1.2 supports per-turn extensions of 3 to 10 seconds, up to 30 seconds in total. You can keep a shot going instead of cutting to a fresh generation that may not match.

Chat editing does not make iteration free. Pricing is per second of output video, and Google's sources don't say how edit turns are billed; until you've tested, assume each turn is billed as new output. Since quick script changes are a known weak spot, changing the spoken line in a clip may mean regenerating more than you expect. A chat history is also not a brief. If five people are giving notes, someone still has to merge them before they reach the model.

What a 60-second ad actually costs

Start with the sticker prices. MarkTechPost puts a full 15-second 720p clip at about $1.05. In MxChat's hands-on test, a 5-second clip cost $0.12 at 360p, $0.35 at 720p, $0.53 at 1080p and $1.05 at 4K. The same 5-second 720p clip cost $0.51 on 1.1.

For higher resolutions, the DEV Community breakdown of Google's pricing page puts upscaled 1080p at about $0.104 per second ($1.56 per 15-second clip) and 4K at about $0.209 per second ($3.13 per clip). The API has no free tier. Google suggests drafting at 360p or 720p and upscaling the winner, in its own post on X. The sources don't say clearly whether the upscale step is billed on top of the draft, so budget as if it is.

The number that matters is cost per approved second, not cost per generated second. If you keep one take in three, your effective 720p rate is $0.21 per second, not $0.07. Drafting at 360p cuts the cost of throwaway takes: in the MxChat numbers, a 5-second 360p test costs about a third of the 720p price.

My view: even with generous retakes, raw generation for a 60-second ad stays in the tens of dollars, not hundreds. The expensive parts are the hours spent briefing, reviewing, fixing the end card and getting sign-off, so put your business-case effort into estimating those. A 60-second spot built from four 15-second clips works out roughly as follows.

• Four 15-second clips at 720p, every take used: about $4.20
• Four clips at the upscaled 1080p rate alone: about $6.24
• Four clips at the upscaled 4K rate alone: about $12.52
• 720p draft plus 1080p upscale, if upscale is billed on top of the draft (unconfirmed): about $1.05 + $1.56 = $2.61 per clip, or $10.44 for four
• 720p draft plus 4K upscale, if upscale is billed on top of the draft (unconfirmed): about $1.05 + $3.13 = $4.18 per clip, or $16.72 for four
• Three 720p takes per shot (12 clips, about $12.60) plus a 1080p upscale of the four winners (about $6.24): about $18.84

Google Vids versus the API for a marketing team

For most marketers, Google Vids is the entry point. Google says you can generate 15-second clips at no additional cost in Vids with a Google account or a Workspace subscription. The Workspace Updates post lists availability on Business Starter, Standard and Plus, Enterprise Starter, Standard and Plus, Education Plus, personal accounts and AI Pro and Ultra users. Clip length is set with a 3-to-15-second slider.

The catch is that base quotas are not published. Users with the AI Expanded Access add-on get higher limits, but the post doesn't list exact numbers for anyone. You can't plan a campaign with dozens of variants in Vids until you have tested your own ceiling. Do that in a quiet week, not the week before launch.

The API is more predictable but not unlimited. Per the changelog summary, a paid Tier 1 project gets 8 requests per minute and 40 per day, and Tier 3 gets 40 per minute and 2,000 per day. Forty a day goes quickly when each shot needs several takes across formats. Teams producing at real volume can look at the Provisioned Throughput option mentioned in Google Cloud's post. In the Gemini app, 15-second creation reaches AI Ultra subscribers first.

Keeping products, faces and logos consistent across clips

Consistency is the problem that sinks multi-shot AI ads, and Google says it has improved. According to The Next Web, faces, clothes, packaging text and brand colours now hold up better across camera angles, with fewer glitches and jump cuts. In Google DeepMind's own comparisons, testers preferred 1.2 over 1.1 for visual consistency 62% of the time. That is a vendor test against the previous version, so treat it as a relative signal, not a guarantee.

References are how you use that improvement. Google says generations stay more consistent when references are used, so characters or products look the same across clips. OrcaRouter reports support for up to six reference images. Build a fixed reference pack before you write a single prompt: front and side shots of the product, a clean packshot, the presenter or character, and a frame showing the colour palette.

Some things I would still not leave to the model. Small packaging text and logos are where generative video slips, and a slightly wrong logo in a paid ad is worse than no logo. Overlay the logo, price and legal lines in the edit. Expect lighting and colour to drift between separately generated clips too, and plan a colour-grading pass on the stitched timeline.

• Fix one reference pack per campaign and version it
• Keep camera language consistent in prompts (lens feel, time of day, palette)
• Add logos, prices and disclaimers in post, not in the prompt
• Review the stitched timeline, not just individual clips

Approvals, rights and the SynthID watermark

Every Omni 1.2 Flash video carries an imperceptible SynthID watermark, as both daily.dev and Crypto Briefing report. Viewers can't see it, but tools that check for it can identify the footage as AI-generated. If your clients, platforms or industry have AI disclosure rules, plan for that from the start.

On commercial use, I have to be blunt: the sources checked for this review did not give specific commercial-use or indemnity terms. Terms may differ between Vids under a Workspace contract, the consumer Gemini app and the paid API on Google Cloud. Have someone read the terms for the surface you actually use before the footage goes into paid media.

The bigger gap is process. A raw model gives you clips, not an audit trail. It won't record who approved the script, which reference images were current or which end card went live in which market. Put a brief and storyboard sign-off in front of generation, in whatever project tool your team already uses, and log the reference pack version with each approved clip. Get approval before generation, not after, or you pay for footage nobody signed off on.

Who should use it now, and who should wait

It is a good fit for performance marketers who need many short variants, e-commerce teams making product clips for social and CTV, and agencies building animatics to sell a concept before a shoot. For those uses, a 15-second limit is not a constraint, and at these prices you can afford to throw takes away.

It is a weak fit for 60-second brand films with recurring characters and dialogue, unless you have an editor who will stitch, grade and fix continuity. It also struggles with fast action and physics-heavy product demos. In regulated categories, where every frame and claim needs legal review, the review cost is likely to dominate whatever you save on generation.

My overall take is that the 30% price cut matters less than the longer clips and the chat edits. Raw generation is already cheap enough that it is rarely the bottleneck. Fifteen seconds isn't the ceiling in this market either: ByteDance's Seedance 2.5 makes clips up to 30 seconds, per daily.dev, while Omni reaches 30 seconds only by extending a 15-second clip over several turns, as OrcaRouter describes. Whichever model you choose, results will depend on your brief, your reference pack and your review process.