Every AI video I make used to end the same way. Just as the scene gets interesting, the clip runs out.
Ten seconds is not a shot. It is a tease. You cannot build a product beat, a dialogue exchange, or a clean push-in with a reveal at the end in ten seconds without cutting it into fragments that never quite breathe.
Gemini Omni 1.2 Flash is Google’s answer to that problem, and it is a more interesting answer than a spec sheet suggests. The headline is 15 seconds. The real story is what changed around those 15 seconds: how you edit them, how they hold a reference, and how much it costs to get one wrong.
Let’s look at it the way a filmmaker would.
What actually changed from Omni 1.1
Google is positioning Omni 1.2 Flash as a multimodal video model—text, image, and video in, video with synchronized audio out—rather than a pure prompt-to-clip generator. Four changes matter for production work.
Native 15-second generation. You can generate and edit clips up to 15 seconds in a single pass, up from 10 seconds on 1.1, with duration selectable from 3 to 15 seconds in one-second increments. It is a 50% increase in native length, not an extension trick.
Multi-turn conversational editing. You can refine the same clip turn by turn—swap a product, change the background, adjust lighting or weather, switch aspect ratio—without regenerating from scratch. Google says this works across the full 15-second clip, not just a still. This is the feature that changes the economics of iteration.
Stronger reference adherence. You can combine character, product, or style images with short motion clips (up to three seconds) and audio cues, and the model holds faces, garments, packaging typography, and brand colors across camera angles with less drift. The AI Studio changelog lists consistency across up to six reference images.
More natural speech and steadier camera moves. Dialogue is described as more fluid with clearer lip-sync and distinct multi-speaker conditioning. Complex continuous moves—FPV flythroughs, dramatic zoom-outs, orbits—show fewer unwanted jump cuts and a more stable scale.
On paper, this is a workflow update disguised as a model update. The question is whether it survives contact with a timeline.
The pricing change is the quiet headline
Google cut the 720p generation price by 30% from Omni 1.1, to $0.07 per second**. A full 15-second 720p clip now costs about **$1.05.
More importantly, the draft-to-upscale workflow is now the intended path: generate fast at 360p or 720p, fix a detail with a follow-up conversational edit instead of re-rolling the whole clip, then upscale only the keeper to 1080p or 4K.
That is a meaningful shift for a one-person studio. The old failure mode was the re-roll tax: one bad detail—a logo warped, a sleeve the wrong color, a background that does not match the next shot—forced a full regeneration. You paid again for a clip you already mostly liked.
If conversational editing actually removes even half of those re-rolls, the effective cost per usable shot drops well below the sticker price.
Where it fits: the “instruction-exact” model
I ran through a 20-test review of the Omni 1.2 Flash video API across text-to-video, image-to-video, reference-to-video, video-edit, refine, and upscale endpoints. The short version is worth repeating:
Omni 1.2 Flash is the most instruction-exact video model I have tested. It is brilliant at explainers, typography, chart animation, and single-change edits to real footage. It is not a cinematic, photoreal-people model.
That is not a criticism. It is a routing decision.
Strong at: flow-diagram explainers with labels, kinetic typography (including non-Latin scripts), chart animation when you provide a finished last frame, and wardrobe, background, or claymation-style edits on real footage.
Weak or unreliable for: cinematic ads and photoreal people from scratch, plotting values without a last frame, refine (switched off at launch), footage with real brand logos due to a copyright filter, and maps with more than four labeled stops.
For Frame by Frame AI readers, the implication is straightforward. Omni 1.2 Flash is not the tool you reach for when you need a hero shot of a human face in dramatic light. It is the tool you reach for when you need a precise, editable, text-bearing, product-accurate shot that cuts into a larger sequence—and when you need to make five variants of it without rebuilding the scene each time.
The conversational editing is the real differentiator against the field. Kling and Runway tend to treat editing as a separate model pass—slower and more expensive. Omni performs the edit inside the same generation loop, so iteration feels closer to a text chat than a render queue.
What it does not solve
A few honest limits, because this site is not a launch post.
The model does not support system instructions, structured output, image generation, or interleaved image-and-text output in the same way a general Gemini model might. It is a video generation and editing model, not a universal creative agent.
1080p and 4K are upscaled, not natively rendered at those resolutions. The native generation resolutions are 360p and 720p, with 1080p/4K as a finishing step.
Complex sports action, rapid script changes, and hard physics remain “active development areas,” according to Google’s own framing.
And the leaderboard story, while real—Omni 1.2 Flash reportedly took first place on both the text-to-video and image-to-video Arena leaderboards and third in video editing—is not the same as a story about whether the footage cuts together in your edit. A leaderboard measures a model’s output in isolation. A timeline measures whether that output has a reason to exist next to the shot before it.
How I would actually use it in a workflow
Here is the routing logic I would apply today, based on the tests and the documented capabilities.
Use Omni 1.2 Flash for: product explainers with on-screen text, catalog spins and virtual try-on from existing product photography, social ad variants that need background or aspect-ratio swaps, kinetic typography sequences, and any shot where reference fidelity matters more than photoreal skin.
Draft at 360p or 720p. Edit in conversation. Change the one thing that is wrong—the product color, the background, the weather, the aspect ratio—rather than regenerating. Upscale the keeper only after the cut is locked.
Do not use it as your primary photoreal human generator. For a close-up of a face carrying an emotional beat, route elsewhere. Omni is a director’s precise instrument for controlled, editable, information-bearing shots. It is not a cinematographer’s replacement for a lit human performance.
The demo looks good. Now let’s see if it survives the edit.
Omni 1.2 Flash has a better chance than most. Not because it makes prettier clips, but because it makes correctable clips—and correction is where most of the real work lives.
A good shot is not just generated. It is chosen. And increasingly, it is revised in conversation.
FAQ
What is Gemini Omni 1.2 Flash?
Gemini Omni 1.2 Flash is Google’s multimodal video generation and editing model, generally available through the Gemini API, Google AI Studio, and Google Cloud. It generates and edits video with synchronized audio up to 15 seconds at 360p, 720p, 1080p, or 4K in a single model.
How long are the clips?
Native generation supports 3 to 15 seconds in one-second steps, up from 10 seconds on Omni 1.1. First-and-last-frame interpolation runs 3 to 10 seconds, and per-turn video extension can build up to 30 seconds total across multiple turns.
How much does it cost?
720p generation is $0.07 per second**, a 30% reduction from Omni 1.1. A 15-second 720p clip costs about **$1.05. Google also offers a draft-to-upscale path: generate cheap at 360p/720p, edit conversationally, then upscale only the final take to 1080p or 4K.
What is multi-turn conversational editing?
It lets you refine the same clip across multiple turns—swap a product, change the background, adjust lighting or weather, switch aspect ratio—without regenerating the entire clip from scratch. Google says it works across the full 15-second clip, which makes it useful for producing ad variants without re-briefing the model each time.
Is it better than Kling or Runway for filmmakers?
It is different, not universally better. Omni 1.2 Flash is strongest when instruction precision and editability matter more than raw cinematic realism. Kling and Runway still have an edge in certain photoreal and multi-shot consistency scenarios. Omni’s conversational editing loop is its clearest workflow advantage for iterative, text-bearing, product-accurate work.
Can I use it in Google Vids or Flow?
Yes. Google says 15-second clips are available in Google Vids at no additional cost with a Google account or Workspace subscription, and conversational editing is rolling out in Flow. Google AI Ultra subscribers get it in the Gemini app.
Does the output carry a watermark?
Yes. Output carries an imperceptible SynthID watermark, and the model supports Content Credentials (C2PA).
What happens to Omni 1.1 Flash API requests?
Effective December 3, 2026, API requests to gemini-omni-1.1-flash will automatically route to gemini-omni-1.2-flash at 1.2 Flash rates.
