How Much To Run AI
Hosted API

Gemini Omni 1.2 Flash: Video Generation, Editing and Cost

Gemini Omni 1.2 Flash generates video with synchronized audio and edits existing clips through follow-up instructions. Google made it generally available on November 5, 2026. Use it for short product scenes, explainers and variations of an existing shot; the Gemini API supports 3–15-second generation, with a shorter limit when supplying both first and last frames.

Content reviewed:

Gemini Omni 1.2 Flash video output cost per 15-second clip: 1,931 tokens a second at 360p gives $0.35, 5,792 at 720p gives $1.04, 8,688 upscaled at 1080p gives $1.56 and 17,376 at 4K gives $3.13; three attempts per accepted 720p clip brings 100 clips to $312.77.
Original diagram of the published token conversion and rates; no clip was generated for it.

Use this model

Online access and trials

API docs and console

Use cases

  • Generate a short scene with dialogue, music or environmental sound.
  • Change a background, product or lighting while retaining an existing performance.
  • Draft at a lower resolution, then upscale a selected clip.

Choose the Gemini API or a creative app

For an application, use the paid Gemini API with model gemini-omni-1.2-flash through the Interactions API. Google Cloud offers a separate enterprise access path. For an interactive edit, Google lists Flow, Google Vids and a rollout to Google AI Ultra subscribers in the Gemini app. App allowances and API usage are separate budgets.

Omni is a hosted video model. There is no downloadable checkpoint or local GPU setup in this release. Its video request belongs on the Interactions API rather than a text-model generateContent or Chat Completions endpoint.

Gemini Omni 1.2 Flash API pricing

The Gemini API Standard rates checked November 6, 2026 are $1.50 per million input tokens for text, images and video, $9 per million text output tokens, and $12 per million video output tokens. The published output schedule includes thinking tokens. There is no free API tier for this model.

Video token consumption changes with resolution. At 720p, 5,792 tokens per second × $12 per million gives $0.069504 per second. A 15-second video therefore has about $1.04 in video output charges. Input and any text output add to that amount. Google describes the 720p price as 30% below Omni 1.1.

Output resolutionVideo tokens / second15-second video output cost
360p1,931$0.35
720p5,792$1.04
1080p, upscaled8,688$1.56
4K, upscaled17,376$3.13

Budget attempts and editing rounds

For 100 accepted 15-second 720p clips, one attempt each gives $104.26 in video output charges. Three generated attempts per accepted clip give $312.77, before input, text output and additional edits. This planning example uses the official token conversion; replace the attempt count with your own acceptance rate.

An edit produces another video output. Include each billed revision and any upscale in the project budget. A channel reporting an uncharged refusal does not establish that every failed or retried request is free across providers.

Match duration to the input mode

The Gemini API guide allows 3–15 seconds in one-second increments, 16:9 or 9:16 output, and 360p or 720p generation with 1080p/4K upscaling. First-and-last-frame interpolation remains limited to 3–10 seconds. A source clip for editing must be 15 seconds or shorter.

For a new generation, the Gemini API supports up to six reference images. An 11–15-second output can use one reference video up to three seconds long; shorter outputs can use up to three such videos. Reference-video sound is ignored, and uploaded audio references are unsupported in the current Gemini API. Google’s launch demonstrations span multiple products, so use the chosen API guide for input limits.

Try a product shot with a clear label and a simple camera move first. Check the label, object shape, hands and final frame before adding multiple actors or fast action. Google still identifies fast sports action and complex physics as areas it is improving.

Generate a first 15-second video

Enable billing for your Gemini API project and install the current google-genai Python SDK. Keep GEMINI_API_KEY in your server or local environment. This example follows Google’s Interactions guide and writes the returned inline video to scene.mp4.

Open the resulting file and check the complete scene, sound and ending. Keep the returned interaction ID if you intend to edit it. For larger videos, the guide recommends URI delivery instead of inline data; use the guide’s download workflow for that response type.

Paid Gemini API; install google-genai and set GEMINI_API_KEY first
import base64
from google import genai

client = genai.Client()
result = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input="One continuous shot of a paper airplane gliding through a quiet library. Soft room ambience, no dialogue.",
    response_format={"type": "video", "duration": "15s", "resolution": "720p"},
)
with open("scene.mp4", "wb") as video:
    video.write(base64.b64decode(result.output_video.data))
print(result.id)

Keep the interaction for a follow-up edit

In the Gemini API, send previous_interaction_id with a short instruction such as “Change the room lighting to dusk. Keep everything else the same.” Setting store=false on the original generation prevents this stateful editing path. Voice editing is unsupported; an extension appends to the end rather than inserting footage in the middle.

In the EEA, Switzerland and the UK, the Gemini API currently permits editing or extending model-generated clips but does not permit those operations on uploaded videos. Confirm the source-video route before planning a workflow around your own footage.

Follow-up using the result from the preceding Python example
edited = client.interactions.create(
    model="gemini-omni-1.2-flash",
    previous_interaction_id=result.id,
    input="Change the room lighting to dusk. Keep everything else the same.",
)
with open("scene-dusk.mp4", "wb") as video:
    video.write(base64.b64decode(edited.output_video.data))

Early tests show differences between channels

Piotr Olszewski’s November 5 Promptowy test reports a 15-second 720p FPV clip through OpenRouter at $1.04 with an 81-second wait. He found the person and setting coherent through that shot. In a separate editing report, all five ways he tried to edit a generated cat clip failed without a charge. Promptowy also promotes a shop selling AI subscriptions.

Segmind’s Himanshu Goel reports a 15-second 720p vertical ad billed at $1.11 and taking 77 seconds. His endpoint table limits reference-to-video and paired-frame generation to ten seconds, while single-frame and text generation reach fifteen. Segmind sells this API access, so these are supplier tests of its own service.

In another Segmind test, Rohit Rao found its refine endpoint unavailable at launch even though video-edit accepted source footage. If follow-up editing is essential, test that exact endpoint before a large batch. These launch-day reports describe their channels and tasks; they do not establish a latency or failure rate for Google’s direct API.

Choose the workflow before comparing video models

Omni’s main decision is whether you want to refine one short shot through a conversation. MiniMax H3 offers a separate open-weight video path; Kling 4.0 has its own reference controls and provider pricing. Match the duration, input mode, resolution and audio before comparing the cost of an accepted clip.

Before you use it

  • Check the exact channel’s duration and editing support before committing a production batch; app and intermediary limits can differ from the direct Gemini API.

How to get started

Start with one 720p scene, save the interaction ID and actual token usage, then budget the revisions needed for an accepted clip.

Sources

Sources (8)

Calculators