Gemini Omni 1.2 Flash: edit, don’t re-roll
Google makes Gemini Omni 1.2 Flash generally available. Clips up to 15 seconds with sound that you rework by text instead of regenerating.
Miraki MindAI-assisted editorial · 15 November 2026 · 7 min read
On 5 November 2026, Google made Gemini Omni 1.2 Flash generally available on Google Cloud. The video model makes clips of up to 15 seconds with sound. It reworks existing clips from a text instruction: swap the product, change the light, switch the camera angle, without starting over. Developers get in through the API, business customers through Google Vids. Gemini for business is supposed to follow soon.
This is exactly the kind of edit we’ve been chewing on for weeks. A clip has a good performance, but one single thing is off. Then the whole clip shouldn’t have to be made again. For that we built our own recipe with Boffjitsu. Nano Banana makes a target frame with the change. Seedance 2.5 then redraws the whole clip and takes the motion from the original. Two models, one chain, a few manual steps. Omni promises the same in one model, in conversation. We haven’t tested it in detail yet. But for fast content and for product adaptations, meaning the same clip for other languages, markets or audiences, we’ll definitely use it.
What’s actually new
Editing itself isn’t new in Omni. Google’s model page already lists video editing for Omni Flash. According to Google, what’s new in 1.2 is mainly length, price and how closely the model sticks to the references.
- Length: generate and edit up to 15 seconds instead of the previous 10.
- Draft, then finish: drafts in 360p or 720p, with the best takes upscaled afterwards to 1080p or 4K.
- References: the model is supposed to stick more closely to image and video references, so products, characters and brand details stay the same from shot to shot. Audio references are only available after approval.
- Editing in conversation: keep refining the same clip instead of restarting, so swap the product, change the light or switch the camera angle.
- Speech: smoother dialogue, cleaner lip sync, multiple speakers clearly separated.
- Price: 0.07 US dollars per second at 720p, 30 percent less than Omni 1.1 Flash. A 15-second clip at 720p therefore costs about one dollar.
How well this holds up in practice isn’t shown by the customer quotes in the blog. Google picked them itself, and they sound like it:
Gemini Omni 1.2 Flash edits exactly what you ask for, a colorway or a background, and leaves the product untouched …
Andrew Taco Zhang, FLORA, in the Google Cloud Blog
What this means for a production
Draft small, finish big: that’s how we already work. With Boffjitsu we test at 480p. Only the run we keep gets re-rendered at 1080p with identical inputs. Omni offers this step as an upscale inside the same model, which saves a pass. Whether the upscaled image then matches the approved draft exactly isn’t something Google’s announcement says.
15 seconds is enough for a whole scene in one go. That means fewer transitions between individual generations, where light or character can tip over.
The real question is the edit. From our Boffjitsu tests we know where this kind of thing breaks. Only one change per run, otherwise it gets shaky. Text on reference images ends up in the video even when the prompt forbids it. We never take the sound from the AI; the original audio goes underneath at the end. And Seedance shifts the timing slightly: without compensation, the measured motion fidelity in our test dropped from 0.85 to 0.29. Whether Omni holds timing and camera movement across several edits in conversation is the test that counts for us. Honestly: an edit that eats a bit more of the original with every round is useless in post.
What this means for a brand
Labelling: According to Google, all outputs carry an invisible SynthID watermark. The documentation lists Content Credentials (C2PA) as supported for the model. That’s the kind of machine-readable marking the EU AI Act requires from providers (more on this). It doesn’t replace the visible notice for deepfakes. That stays the job of whoever publishes the film. And whether the marking survives cutting, grading and encoding is still open.
Data: Omni runs in Google’s cloud. Anyone who wants to edit a shot clip by text uploads it there, with everything that’s in it: a product before launch, faces from the cast. With us, what runs locally and what runs in the cloud is decided together for each project. Scripts, faces and product data stay with us unless agreed otherwise. A cloud edit therefore has to be agreed beforehand.
Switchover: From 3 December 2026, the Gemini API automatically redirects all requests for Omni 1.1 Flash to 1.2, billed at the 1.2 rate. Anyone who automates product videos through this API will then get a different model without changing a line of code. It’s worth checking look and consistency beforehand.
Opportunity: Google is explicitly building the model for videos that companies make in bulk: product videos, virtual try-ons, ads. Anyone who needs many variants of the same clip, per colour, per market or per language, is exactly who it’s meant for. That’s exactly where we want to use it: adapting an approved clip for another language or another audience without rebuilding the scene.
Our take and the limits
Omni 1.2 Flash isn’t a leap but an everyday version: longer, cheaper, closer to the references, with upscale in the same model. For brand work, the edit in conversation is the interesting part. It’s closer to the logic of a post-production than rolling the dice on new takes. For fast content and adaptations, we’ll definitely add the model to the toolbox. Does it also bring the control we force through several steps with Boffjitsu? We’ll only know after a test on our own material.
Quite a bit remains open. How stable do timing and camera stay across several edits? How well does lip sync hold when the same clip gets another language? What exactly do upscale and the other resolutions cost, where Google only mentions «similar savings»? And what happens to SynthID and C2PA when an Omni clip meets shot footage in the edit?
Sources
- Google Cloud Blog: Gemini Omni 1.2 Flash now GA, 5 November 2026
- Google Cloud documentation: Gemini Omni 1.2 Flash, accessed 14 November 2026
- Google AI Studio: Gemini Developer API pricing, accessed 14 November 2026
- Google DeepMind: Gemini Omni, model page, accessed 14 November 2026
- Miraki Mind: Boffjitsu, video-to-video, the performance stays, 1 November 2026
- Miraki Mind: AI labelling, what changes on 2 December, 26 October 2026