Gemini Omni 1.2 Flash: Create and Edit 15-Second AI Video, 30% Cheaper

Why does 10-second AI video always cut your story short?

It is 9 p.m. and the launch ad has to be ready before tomorrow morning. You already have a studio photo of the product and an approved tagline. Last week’s 10-second AI video only gets as far as the opening: the model turns the product, the camera orbits once, and the clip ends right before the tagline appears. Want five more seconds to state the product’s value? Not possible. Want to change the model’s shirt colour to match the new brand identity? You cannot edit it — the only option is to regenerate the whole clip, and that gives you a different face, different lighting, and a different camera angle. Five regenerations mean five different clips, five payments, and none keeps what you liked about the first one.

That is the pain Gemini Omni 1.2 Flash is built for. Google took this model to GA on Nov 5, 2026 on the Gemini API and Gemini for business: clips up to 15 seconds long, editable in place through multi-turn conversation, and 720p pricing down to $0.07/second.

What is Gemini Omni 1.2 Flash?

How is Gemini Omni 1.2 Flash different from an AI video app?

First, one clarification: Gemini Omni 1.2 Flash is not an app. It is a video model from Google DeepMind, and the first model in the Gemini Omni line to leave preview. You reach it through several doors: the Gemini API and Google AI Studio if you write code, the Google Cloud console for enterprise solutions, Flow if you direct through an interface, Google Vids for presentation videos, the Gemini app on the Ultra plan, and Google Ads for advertising campaigns.

The best fit is developers integrating video into a content pipeline, creators and marketers who need short ad variants, and DevOps or IT managers who need a clear cost figure to approve a budget. If you need videos several minutes long, complex scripts, or a model that runs locally on a homelab, this is not the right choice.

The three biggest changes from Omni 1.1

Fifteen seconds looks like just five extra seconds, but in my view the real trio is length plus multi-turn editing plus reference following: length fits a complete short ad story, multi-turn editing changes product, background, lighting, weather, or aspect ratio without starting over, and reference following keeps faces, clothing, packaging typography, and brand colours stable across angles.

Architecture and how Gemini Omni 1.2 Flash works

The main components in the AI video pipeline

Input consists of a text prompt, up to 6 reference images, and — depending on clip length — a motion reference clip of up to 3 seconds: clips of 11–15 seconds accept 1 reference clip, clips of 3–10 seconds accept up to 3. Google’s documentation also lists audio cues as an input type, but note this right away: in the current API, audio input upload and audio reference are not supported; sound is generated by the model from the prompt, for example background music or dialogue described in words.

At the core is the Omni 1.2 Flash engine, with multi-reference conditioning and temporal consistency: characters, objects, and scenes are held from the first frame to the last, with fewer jump cuts. The model also handles multi-speaker dialogue with per-character lip-sync. Default output is 720p; you can choose 360p for cheap drafts, or 1080p and 4K for the final version. Every clip carries an imperceptible SynthID watermark for provenance tracing.

Data flow and control flow: draft first, then upscale

flowchart TD
    A["Input<br/>Text prompt + Reference images<br/>character / product / style<br/>Motion clip ≤ 3 seconds<br/>Audio cues + First/last keyframe"] --> B["Access point<br/>Gemini API / AI Studio<br/>Flow / Google Vids"]
    B --> C["Gemini Omni 1.2 Flash engine<br/>Multi-turn session + Reference following<br/>Temporal consistency"]
    C --> D["Draft clip 360p / 720p"]
    D --> E{"Needs edits?"}
    E -->|"Yes — change product, background,<br/>lighting, aspect ratio, camera angle"| C
    E -->|"No — approved version"| F["Upscale the chosen version<br/>to 1080p / 4K"]
    F --> G["SynthID watermark"]
    G --> H["Publish<br/>Gemini app / Google Ads<br/>Your own channel"]

The key point is where the edit loop sits: it returns to the engine before upscaling, not after. You draft at 360p or 720p, edit until you are happy, and only then upscale the chosen version. Cost and quality are different levers: drafts are cheap for testing ideas, upscaling is only for the winner.

A practical note about “upscale” via the API: the documentation describes no separate upscale endpoint. Resolution is chosen in response_format when the clip is generated. So the “draft then upscale” workflow via the API really means regenerating the winning version at 1080p or 4K from the same prompt and references, while in Flow upscaling is an in-app action. Do not go looking for an upscale API that does not exist.

Known dependencies and technical limits

To use the API, you need a Google account, a Google Cloud project with billing enabled, and a paid API key — this model has no free tier. Through Flow or Vids you use a Google or Google Workspace account, and some advanced features in Flow follow the Google AI plan; before paying, check Flow’s plan page instead of trusting any secondary pricing table.

Google itself admits three limits still being improved: fast sports action, rapid script changes, and complex physics. The API documentation adds: first/last keyframe interpolation only supports clips of 3–10 seconds; editing or extending self-uploaded video is unavailable in the European Economic Area, Switzerland, and the United Kingdom; voice editing, system instruction, temperature, and negative prompt are unsupported — to exclude something, write it into the normal prompt, for example “No dialogue”. Setting store=false also makes the generated clip uneditable in later turns.

How to use Gemini Omni 1.2 Flash: from a quick test to API calls

Step 0 — Preparation: account, access, and expected cost

Prepare three things: an API key from a project with billing enabled (created in Google AI Studio), sharp reference images of the product or character, and an estimated cost table. Pricing is per output token: $12 per million video tokens, with 720p costing 5,792 tokens per second, or about $0.07/second. Google’s documentation gives ready conversions: a 10-second clip costs about $0.70, a 15-second clip about $1.04; 360p about $0.023/second, 1080p about $0.104/second, and 4K about $0.209/second. Input tokens cost an extra $1.50 per million tokens, applied to text, images, and video inputs together.

These are published prices, excluding edit turns: each edit turn generates a new full-length clip, so a 4-turn round on a 15-second 720p clip costs about $4.16 in output alone, excluding input.

Method 1 — Quick no-code test with Flow and Google Vids

If you only need to test an idea or make a single clip, Flow is the shortest path: enter a prompt, direct the scene, edit multi-turn with natural-language commands, use Extend to lengthen it, then upscale the best version. Google Vids creates 15-second clips at no extra cost with a Google or Google Workspace account, which suits presentation videos or feature-introduction clips. The Gemini app also creates 15-second videos, currently rolling out to Google AI Ultra users.

This suits a team demo or validating an idea before writing code. The downside is that you cannot control fine-grained parameters such as per-second duration or the interaction ID used to chain edit turns in an automated pipeline.

Method 2 — Calling the Gemini API via AI Studio

You need Python 3.10 or later. Install the SDK and set the key via an environment variable — never hardcode the key into code:

pip install -U google-genai
export GEMINI_API_KEY="YOUR_API_KEY"

Check that the SDK picked up the right package:

python -c "from google import genai; print('google-genai OK')"

Expected result: the line google-genai OK. If you see ModuleNotFoundError, check that the virtual environment in use is the one where you just ran pip install.

Create a 5-second 720p draft clip and save it as an MP4 file. The model ID is gemini-omni-1.2-flash, called via the Interactions API (POST /v1beta/interactions), not generateContent as with text models:

import base64
from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from the environment

draft = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input="A ceramic coffee cup on a wooden table, morning light, "
          "slow orbit shot, no dialogue.",
    response_format={
        "type": "video",
        "duration": "5s",       # "3s" to "15s", default "10s"
        "aspect_ratio": "16:9",  # or "9:16" for vertical clips
        "resolution": "720p",    # 360p / 720p / 1080p / 4K
    },
)

with open("draft.mp4", "wb") as f:
    f.write(base64.b64decode(draft.output_video.data))

print("Interaction ID:", draft.id)

Check whether the file was generated:

ls -lh draft.mp4

Expected result: a draft.mp4 file of a few MB that opens in a regular video player. The returned video is base64 in interaction.output_video.data when using the SDK; with a direct REST call, it sits in the steps array under model_output. For clips larger than 4 MB, set delivery="uri" in response_format to receive a link instead of a base64 payload.

Edit the clip in the next conversation turn, using previous_interaction_id to point back to the previous turn instead of re-uploading the video:

edited = client.interactions.create(
    model="gemini-omni-1.2-flash",
    previous_interaction_id=draft.id,
    input="Change the cup color to deep green. Keep everything else the same.",
)

with open("edited.mp4", "wb") as f:
    f.write(base64.b64decode(edited.output_video.data))

Keep edit prompts short and specific; the documentation recommends adding “Keep everything else the same” when you only want to change one detail, because a wordier edit prompt is more likely to change parts you never mentioned. To lengthen a clip, use a prompt like “Extend this video” — each extend turn adds 3–10 seconds, up to 30 seconds in total, and the model only uses the last 10 seconds of the source clip as context.

Common errors and how to fix them

There is no undo button here. The only rollback is to return to the best version you saved: keep the prompt, reference images, and interaction ID of every turn you liked, save the draft MP4 after each good turn, and when an edit drifts the scene, regenerate from that good version’s interaction ID. Without snapshots, you have no point to return to except starting over — exactly what multi-turn editing exists to avoid.

Technical review of Gemini Omni 1.2 Flash

Pros

A 15-second clip is enough for a complete short ad, and 1-second steps let you price cost accurately to the second instead of paying in fixed tiers. Multi-turn editing removes the most expensive regeneration step in the old workflow, while reference following keeps brand identity across variants. At $0.07/second it is 30% cheaper than 1.1, and drafting cheap then upscaling only the winner is a real cost lever.

Cons and open questions

It is still short-form video: at most 15 seconds per generation, 30 seconds after extending — no long videos or fast-changing scripts. Google publishes no benchmark, so every claim about consistency, lip-sync, or a “cinematic look” is a vendor claim without independent verification — this article states it in exactly that tone, without exaggerating further. Edit turns have no separate price table: each turn is billed like a new clip generation, so the real budget for a campaign with many variants is far higher than the $1.04/clip figure. And the whole workflow depends on Google infrastructure: no self-hosting, no local run — from Dec 3, 2026, requests to 1.1 are routed to 1.2, so old code still runs but the bill follows the new pricing.

Security and responsibility when using AI video

Every clip carries a SynthID watermark, which helps trace provenance, but a watermark does not replace content review: any clip using images of real people or real products needs image-use rights and an approval step before publication, to avoid deepfakes or impersonation. Keep the API key in an environment variable or a secret manager — never hardcoded, never committed to a repo; set quotas and cost alerts so a broken edit loop cannot burn the whole budget. Log the prompt, reference images, and interaction ID of every published clip — the only trail for investigating a clip if a complaint comes in.

Performance, scalability, and operations

Only verified figures are stated here: $0.07/second at 720p, 3–15 seconds in 1-second steps, 1080p/4K upscaling, and extension up to 30 seconds. The documentation publishes no generation speed or success rate, so this article gives no figures for either. For teams, the process should be draft, review, upscale, then publish, with a cap on edit turns per clip and a per-second budget per campaign. If your team is on Omni 1.1, prepare the migration before Dec 3, 2026: change the model ID, rerun the test prompt set, and recheck the invoice, because pricing follows 1.2.

Gemini Omni 1.2 Flash vs Veo 3.1, Sora, and Runway

The table reveals something few people notice: in short-form video via API, the two biggest names of last year have both left the game. Veo 3.1 Preview was retired in the Gemini API from Oct 22, 2026, with the official documentation recommending migration to Omni; Sora 2 is likewise no longer a practical API option after Sep 24, 2026. The meaningful comparison left is Omni 1.2 versus Runway Gen-4.5: Omni wins on longer clips (15 vs 10 seconds), multi-turn editing in the same session, and a lower per-second price; Runway wins on its long-established app ecosystem and monthly-plan credit model. In my view, if your workflow is code and automated pipelines, Omni 1.2 is the clear choice right now.

Real-world uses for 15-second AI video

Business and marketing: fast ad variants

The clearest use case is generating many ad variants from the same product image: change the background by season, shift lighting from day to night, switch the aspect ratio from 16:9 to 9:16 for each placement, while the model’s face, packaging, and brand colours stay consistent thanks to reference following. Draft at 360p for fast review, and upscale only the winner for A/B testing. At about $0.35 for a 5-second 720p draft, testing an idea costs far less than a reshoot.

Creators and small content teams: product demos without filming

A small team can make product demos, feature clips, or social videos from images and prompts, with no filming set. Inspect hands, movement, and lip-sync closely before posting — those are where AI video most often fails.

AI workflows and DevOps: automating a video pipeline with the API

For technical teams, Omni 1.2 plugs into a content pipeline: a fixed prompt template plus fixed reference images, generate, review, upscale, then push to a channel. Cost is controlled with three levers: length in 1-second steps, a cap on edit turns, and project quotas. To be clear: this is automation via API, not self-hosting — the model runs on Google infrastructure, and your homelab cannot install it. The spirit of using AI to replace one manual step is the same as in Craft Anything rebuilds software with AI, only the output here is video instead of code.

When NOT to use it

Do not use it for long videos, fast sports action, complex physics, or rapidly changing scripts — Google itself admits these limits. Do not use it when a project requires an independently verified quality benchmark, because Google publishes none. Do not use it for local or self-hosted workflows, or with images of real people without image-use rights.

Conclusion: who should use Gemini Omni 1.2 Flash now, and who should wait?

Use it now if you are a marketer or creator who needs short ad clips with fast edits, a developer who needs a per-second-priced video API for a pipeline, or a team already using Google AI, Flow, or Vids. Wait or reconsider if you need long videos, independently verified cinematic quality, a local or self-hosted workflow, or if you are on Omni 1.1 or Veo 3.1 and have not yet verified the December 2026 migration path.

Next step: try a 5-second clip in Flow or Vids, then call a draft via AI Studio with your real project’s prompt and reference images, and price one complete variant — including edit turns — before committing a campaign budget.

Lessons learned

1. Fifteen seconds is only valuable together with multi-turn editing: length alone does not solve the review-and-edit problem, it only moves the cut-off from second 10 to second 15.
2. Draft cheap, then upscale, is the only way to keep costs low when pricing is per second: review at 360p/720p, and pay for 1080p/4K only for the winner.
3. Snapshot the prompt, reference images, and interaction ID after every good turn, because session state can drift and store=false locks out further editing entirely.
4. Do not promise quality beyond the limits Google has admitted: fast sports action, rapidly changing scripts, and complex physics are still being improved.
5. Verify the endpoint, edit-turn pricing, and migration date against official documentation before moving to production: the model ID and the Dec 3, 2026 reroute date are verified, but Flow plan pricing and regional limits still need checking for your specific project.

Verification scope

Check date: Nov 9, 2026. Verified against Google’s official launch blog (GA Nov 5, 2026, clips of 3–15 seconds in 1-second steps, $0.07/second for 720p, multi-turn editing, reference following, SynthID, access channels) and the Gemini API documentation in AI Studio (model ID gemini-omni-1.2-flash, Interactions API POST /v1beta/interactions, previous_interaction_id, token pricing of $1.50 input and $12 video output per million tokens, no free tier, Veo 3.1 Preview retired from Oct 22, 2026, 1.1 rerouted to 1.2 from Dec 3, 2026). Defaults: 10 seconds, 720p, 16:9. Price figures are published token-based prices, not measured benchmarks; Google publishes no quality benchmark. Not tested hands-on on a real account — the code is copied from official documentation examples and was not executed by the author — and figures on Sora 2, Runway Gen-4.5, and Flow plan limits come from cross-checked secondary sources, so recheck them before making budget decisions.

References

• Google — Gemini Omni 1.2 Flash: 15-second video and lower pricing
• Google AI Studio — Generate and edit videos with Gemini Omni Flash
• Google AI Studio — Gemini Developer API pricing
• Google AI Studio — Video generation in the Gemini API