Gemini Omni 1.2 Flash: edit a clip by talking to it

Google made Gemini Omni 1.2 Flash generally available on November 5, 2026, and the feature that changes a working editor's day is multi-turn conversational editing. You refine the same clip turn by turn, swapping a product or a background, changing the lighting or the weather, or switching the aspect ratio, instead of re-briefing and regenerating from scratch. The model, which Google describes in its launch post, keeps the scene in memory across turns, so you describe only the change. An edit is now a sentence, not a re-render.

The developer documentation calls the feature conversational editing and extension, and runs it through the Interactions API. Each turn produces a new video, and the model reads context from the turns before it.

What changed for someone editing

The old loop was expensive in time. You wrote a prompt, waited, watched the result, and if the jacket color was wrong you rewrote the prompt and generated the whole clip again from scratch. Gemini Omni 1.2 Flash changes the unit of work. You keep the take you have and direct the change on top of it.

The cost model follows. Google lists Gemini Omni 1.2 Flash at $0.07 per second of 720p video, 30% less than Omni 1.1 Flash, which the pricing page puts at about $0.10 a second. The page bills on output tokens: 5,792 tokens for every second of 720p, at $12.00 per million output tokens. That arithmetic gives roughly $0.70 for a 10-second clip and $1.04 for a 15-second one. A conversational fix runs a new 15-second generation, so it lands at about the same $1.04, and you stop paying the re-briefing time that used to sit on top of it.

How to edit a clip by talking to it

The workflow has four moves, and the order is what keeps the bill down.

Start a draft. Generate at 360p or 720p, the two draft resolutions Google's Cloud post names. A 360p second costs $0.023, so a 15-second draft is about 35 cents. This is the stage to be loose with, because you are looking for the take, not the finish.

Keep the take. When a clip works, do not re-describe it. The documentation says each turn understands context from prior turns, "letting you make incremental changes like adjusting lighting, and swapping backgrounds, without re-describing the entire scene." You describe the delta only.

Direct the change. This is the sentence. Google's launch post lists the moves you can ask for: swapping a product or a background, changing the lighting or the weather, or switching the aspect ratio. The documentation adds the setting: aspect_ratio takes "9:16" for portrait or "16:9" for landscape, with landscape the default. So a 16:9 ad becomes a 9:16 cut for a phone feed without a new brief.

Upscale the keeper. Once the clip is right, move it up. The launch post says you upscale your best creations to 1080p or 4K. The pricing page puts 1080p with upsampling at $0.104 a second and 4K with upsampling at $0.209 a second, so a 15-second 4K finish is about $3.14.

Where each step happens

The same model sits behind three doors, and which one you open depends on who you are.

Developers reach it through the Gemini API under the model ID gemini-omni-1.2-flash, which the developer documentation uses in every example. That door is on the paid tier; the pricing page states the model is "now generally available to developers on the paid tier of the Gemini API," and its free-tier column reads "Not available."

Creators who want to direct scenes and make multi-turn edits without code work in Flow, Google's browser studio. The Flow page lists Gemini Omni Flash among the features included without a Google AI subscription, which brings 50 daily Flow credits. Upscaling steps up by plan: 1080p video upscaling arrives with Google AI Plus at $4.99 a month, and 4K image and video upscaling with Google AI Ultra at $99.99 a month.

Teams and office users get 15-second clips in Google Vids. The launch post says Vids generates 15-second clips at no additional cost with a Google account or a Google Workspace subscription. The Gemini app is a fourth door, and the launch post says 15-second video there is rolling out to Google AI Ultra subscribers.

What a reference controls now

A reference is the material you hand the model and ask it to hold onto. Google's launch post lists the reference types: character, product or style images; short motion clips up to three seconds; and audio cues. The developer documentation sets the ceilings: up to 6 reference images, and reference videos of up to three seconds each (up to 3 when the clip runs 3 to 10 seconds, one when it runs 11 to 15 seconds).

For a product ad, that combination is the useful part. Google's launch post describes a 15-second virtual try-on sequence its shared clips show: one model photo plus front and back product shots of three outfits, generated across 360-degree turns and match cuts, with the model's identity and each garment's cut, drape and back details held consistent.

Audio is the gate. John Nack, a senior product manager at Google Cloud, writes in the Cloud post that "Audio references are gated behind allowlisting." Until an account is allowlisted, treat voice and audio cues as unavailable rather than as a setting you can flip.

The limits worth knowing before you start

Fifteen seconds is the ceiling on a single generation. That is the whole native clip length, and it is also the length cap when you edit a video you upload through the Files API.

The transition window is short. The documentation says gemini-omni-1.1-flash remains available during a four-week transition window, and that effective December 3, 2026, all API requests to 1.1 route to 1.2 automatically and bill at the lower 1.2 rates.

First-and-last-frame interpolation is narrower than single-image generation. If you hand the model a start frame and an end frame, the clip must run 3 to 10 seconds; the documentation says the API returns an invalid argument error when you set a duration over 10 seconds with two frames, while single-image generation still reaches 15 seconds.

Editing a video you upload is not available in every region, the documentation notes. And the model is not on the API's free tier, so a developer without billing cannot try it through that door.

When talking to the clip is the wrong tool

Conversational editing is good at changes to a scene that already works: a background, a color, a light, an aspect ratio, a small product swap. It is a poor fit when the change is structural. If the camera move is wrong or the whole concept misses, no follow-up sentence rescues the take, and you are better off generating a new clip from a new brief.

The same holds for anything longer than the clip. A 15-second generation is a shot or a short ad, not a scene sequence. When a cut runs past that, the answer is an editor, not the model: assemble the shots you have in software built to trim and join them, and spend your generation budget on the pieces you still need.

Google's own note points at where the model still struggles. The launch post says development continues on fast sports action, rapid script changes and complex physics, which is a fair list of what to test before you commit a production to it.

Gemini Omni 1.2 Flash is available now through the Gemini API on the paid tier, in Flow without a Google AI subscription, and in Google Vids with a Google account or Google Workspace subscription; the Gemini app rollout is going to Google AI Ultra subscribers.