Gemini Omni 1.2 Flash Is Here: Generate and Edit 15-Second Videos

Fifteen seconds is the length that pays the bills in short-form video. It is a full Reel or Shorts ad slot: long enough for a hook, a product moment and a payoff. Every Gemini Omni model I tested before this one stopped at ten seconds, so a 15-second spot meant two generations and a stitch in the edit. Gemini Omni 1.2 Flash removes that step. Text-to-video and image-to-video now run from 3 to 15 seconds, and the edit endpoint took a full 15-second clip and handed back 15 seconds.

Interest is clearly there. In Google Trends over the last 90 days, "gemini omni 1.1 flash" was the fastest-rising query attached to Gemini Omni, ahead of MiniMax H3 and Seedance 2.5. People are watching this model family release by release.

I spent an afternoon on the four new Gemini Omni 1.2 Flash endpoints on Segmind: 17 generations, plus one Gemini Omni 1.1 clip on the same prompt as a control. Below are the clips, the exact prompts, what each one cost me, and the limits I hit along the way.

One note before the numbers. As I write this, Google has not published release notes for 1.2, so nothing here is copied from a vendor spec sheet. Every limit is what the API accepted or rejected, and every price is what Segmind billed me per clip on November 5, 2026.

What is new in Gemini Omni 1.2 Flash

Gemini Omni 1.2 Flash is the next release in Google's Gemini Omni video family, and on Segmind it ships as four task-specific endpoints instead of one catch-all. The headline is duration, but the price moved too. Here is how it lines up against 1.1, measured the same way:

Every clip came back as H.264 at 24 fps with a 48 kHz stereo AAC track. The four endpoints:

Use case 1: Marketing agencies, a 15-second vertical ad and its colour variant

This is the use case the extra five seconds were made for. I asked for a three-shot vertical spot in one generation: the lace-up, the sprint, the cooldown.

Scene detection puts the cuts at 8.1 s and 11.8 s, so the model built exactly the three shots I described and gave the opening lace-up and sprint most of the running time. The wet-pavement splash and the backlit skyline both read as a real sunrise shoot. It came back as one finished 15-second file with nothing to stitch.

One lesson cost me a re-render. My first take said only "bright orange sneakers", and the model put a side-stripe pattern on them that looked a lot like a real sportswear brand's mark. Adding "no logos, no stripes and no brand marks" fixed it on the next fire. If a clip is going into a paid campaign, close the label set in the prompt every time.

The variant: same ad, different colourway

Agencies rarely ship one version. I sent that 15-second clip straight into the edit endpoint with a one-line instruction.

The shoes changed and, as far as I can tell frame by frame, nothing else did: same runner, same cuts, same light, same splash. That is a colourway variant for $1.25 without re-shooting or re-prompting the whole spot, and the edit kept the full 15 seconds. On 1.1 the edit endpoint wanted clips of ten seconds or less.

import requests

resp = requests.post(
    "https://api.segmind.com/v1/gemini-omni-1.2-flash-video-edit",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "video": "https://your-bucket.s3.amazonaws.com/ad-15s.mp4",
        "prompt": "Change the running shoes to electric blue. Keep everything else the same.",
        "resolution": "720p",
        "seed": 42,
    },
)
resp.raise_for_status()
open("ad-blue.mp4", "wb").write(resp.content)

Use case 2: Film studios, a 15-second continuous shot and a full-length restyle

For previs, the question is whether a model can hold a planned camera move for the whole shot. I wrote this prompt like a shot list, with timecodes.

It hit all three beats roughly on their timecodes: the wok close-up, the slow pull-back through the steam, and the customer walking off with the plate as the camera lifts. Scene detection found zero cuts, so I asked for a single take and got one. The cook and the customer stay the same people for the full 15 seconds. The weak spot is background crowds: faces at the back soften once the frame goes wide.

Then I restyled the entire 15-second shot.

Staging, timing and camera move all survived. I correlated frames against the source: 0.92 to 0.95 through the first half, dropping to 0.61 to 0.64 in the wide crowd section, where the restyle redraws every face and invents its own stall signage. For a studio, that makes a cheap look-development pass on a real shot. It is the same edit endpoint as the colourway swap, so one integration covers both.

import requests

resp = requests.post(
    "https://api.segmind.com/v1/gemini-omni-1.2-flash-text-to-video",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "A single continuous 15-second shot at a night market. 0 to 5 seconds: ...",
        "duration": 15,
        "aspect_ratio": "16:9",
        "resolution": "720p",
        "seed": 42,
    },
)
resp.raise_for_status()
open("previs.mp4", "wb").write(resp.content)

Use case 3: Production houses and MCNs, a consistent host and an animated product shot

Channel networks need the same face and the same product in every episode. That is what the reference endpoint is for. I gave it two stills: a host and a product.

Reference image 1
reference_images[0]: the host

Reference image 2
reference_images[1]: the product

Both stills were made on Nano Banana 2 and passed in reference_images.

The face, glasses, curls and apron match the reference, and so does the cobalt drip glaze on the vase. I ran the audio through a speech recogniser and the line came back word for word, "Every drip is different. That is the whole point.", between 6.5 s and 9.5 s, which is exactly where she turns to the lens. The catch: reference mode still tops out at 10 seconds, so a 15-second host segment needs two clips.

Animating a product shot between two frames

Image-to-video takes a first frame and, optionally, a last frame. I gave it a moody studio still and the same bottle in full bloom.

First frame
first_frame_image

Last frame
last_frame_image

Input frames for the image-to-video endpoint, made on Nano Banana 2.

Both ends are real locks. The opening frame correlates 0.996 with my first still and the closing frame 0.986 with the last. What surprised me is the pacing. Measured frame to frame, nearly all the motion happens between seconds 1 and 4, and the last five seconds drift slowly towards the end frame. My first try asked for 15 seconds, and the API refused it: with a last frame set, the limit drops to 10. For a transition like this I would ask for 5 or 6 seconds and keep the energy.

Drop the last frame and you get the full 15 seconds back:

The bottle never warps, which is the thing product teams care about most. The "slow orbit" I asked for came out as a gentle drift rather than a real arc, so spell out camera moves if you need them.

Gemini Omni 1.2 Flash pricing, measured

Every figure here is the cost Segmind reported on that exact call. At 720p, cost grows in a straight line with duration, at roughly $0.074 per second:

The other endpoints price the same way. A 10-second reference clip cost $0.748 and a 10-second image-to-video clip $0.745. A 15-second edit cost $1.25, which is the 15-second output plus about $0.13 for reading the source video.

Resolution: draft at 720p, not 360p

I rendered one 5-second prompt at every tier with the same seed:

Then I pulled frames at four timestamps and correlated them. 720p, 1080p and 4K scored 0.999 to 1.000 against each other: they are the same render delivered at three sizes. 360p scored only 0.69 to 0.76 against 720p, so it is a different take of the scene. In practice, a 360p draft does not preview your final. A 720p draft previews your 1080p and 4K exactly, so iterate at 720p, then pay for 1080p or 4K only when the delivery spec demands the pixels.

Seeds hold too. Two image-to-video calls with identical inputs and seed 42 produced frames that correlate 1.000 at every timestamp I checked, and they billed to the same fraction of a cent. Change the prompt and the same seed gives you a different take.

Gemini Omni 1.2 Flash vs 1.1 on the same prompt

Same night-market prompt, same seed, 10 seconds (1.1's maximum), 720p.

Both versions held a continuous take. 1.1 opened tighter on the flames and pulled out to a wide crowd shot; 1.2 Flash framed the cook and stayed closer, landing the plate hand-off clearly. On one prompt I cannot call a quality winner, and I would not trust anyone who claims one from a single pair. What I can say is that 1.2 Flash did the same job for 30.7% less money, nine seconds faster, and can go five seconds longer when you need it.

Developer integration guide

All four endpoints are synchronous: you POST JSON and the response body is the MP4.

import requests

API_KEY = "YOUR_API_KEY"

def omni(slug, **params):
    r = requests.post(
        f"https://api.segmind.com/v1/{slug}",
        headers={"x-api-key": API_KEY},
        json=params,
        timeout=600,
    )
    if r.status_code != 200:
        raise RuntimeError(f"{r.status_code}: {r.text[:300]}")
    print("billed:", r.headers.get("x-cost"))
    return r.content

clip = omni(
    "gemini-omni-1.2-flash-image-to-video",
    prompt="The camera holds on the bottle, then sunlight sweeps across it.",
    first_frame_image="https://your-bucket.s3.amazonaws.com/first.jpg",
    duration=15,
    aspect_ratio="16:9",
    resolution="720p",
    seed=42,
)
open("product.mp4", "wb").write(clip)

The parameters that matter: duration is an integer from 3 to 15, capped at 10 for reference mode or when last_frame_image is set. resolution is 360p, 720p (the default), 1080p or 4k. aspect_ratio is 16:9 or 9:16. The edit endpoint rejects both duration and aspect_ratio, because it follows the source clip. Leave everything out and you get 5 seconds of 16:9 at 720p.

Three gotchas from my run. First, an input URL the API cannot fetch comes back as HTTP 406, the same code as a low balance, so read the error body before you top up. Second, unknown parameters are ignored, not rejected: a stray field will not save you money, and audio is always on. Third, I hit one HTTP 429 ("temporarily unavailable, you have not been charged") that cleared on a retry a minute later. Renders took 23 to 152 seconds, so give your HTTP client a generous timeout or queue the calls. Full parameter docs: Gemini Omni 1.2 Flash on Segmind.

Honest assessment

What it does very well. Fifteen seconds in one generation, with shot structure that follows your prompt, and edits that keep the full 15 seconds. The colourway swap left everything else alone, which is the hardest part of video editing. It is also 30.7% cheaper than 1.1 for the same 10-second clip.

Where it falls short. Reference mode and first-plus-last-frame mode are still capped at 10 seconds, so the 15-second headline applies to text, single-frame image and edit only. First-and-last-frame clips front-load their motion. Background faces in wide crowd shots soften, and the model will invent brand-like details unless you forbid them. 1080p and 4K add pixels, not detail.

Best fit: paid social, ad variants, previs and look development. Not a fit: anything that needs one unbroken generation longer than 15 seconds, or crisp faces deep in a crowd.

FAQ

What is Gemini Omni 1.2 Flash?

Gemini Omni 1.2 Flash is the latest Gemini Omni video model. On Segmind it has four endpoints: text-to-video, image-to-video, reference-to-video and video edit. It generates clips up to 15 seconds with native audio, at up to 4K.

How long can Gemini Omni 1.2 Flash videos be?

3 to 15 seconds for text-to-video and first-frame image-to-video. Reference-to-video and first-plus-last-frame mode cap at 10 seconds. The edit endpoint accepted a 15-second clip and returned 15 seconds.

How much does Gemini Omni 1.2 Flash cost?

In my tests at 720p: about $0.074 per second, so $0.37 for 5 seconds and $1.11 for 15. 360p was about a third of that, 1080p 1.49x and 4K 2.97x.

Can Gemini Omni 1.2 Flash edit an existing video?

Yes. Send a public MP4 URL and a one-line instruction. It kept duration and aspect ratio, changed what I asked for, and left the rest alone. A 15-second edit cost me $1.25.

Is Gemini Omni 1.2 Flash better than 1.1?

It is longer and cheaper: 15 seconds vs 10, and 30.7% less for the same 10-second clip in my tests. On a matched prompt I could not call a clear quality winner.

Is 4K worth it on Gemini Omni 1.2 Flash?

Only if you need the pixels. My 720p, 1080p and 4K frames correlated at 0.999 or higher, so they are one render at three sizes. Draft at 720p and upscale at the end.

Conclusion

Gemini Omni 1.2 Flash makes the 15-second clip a single API call. For agencies, that is a finished vertical ad plus colour variants from the edit endpoint. For studios, it is a full continuous previs shot and a restyle that keeps the move. For channel networks, it is a consistent host who says the line you wrote, and product stills that come to life. Draft at 720p, close the label set in your prompts, and keep first-and-last-frame transitions short.

Try Gemini Omni 1.2 Flash on Segmind. It is available via API with no setup required.