Gemini Omni 1.1 Flash: 40s Video, 4K Upscale

The problem with Google’s video models for the past year was length. An 8-10 second clip is impressive but it isn’t a scene. Gemini Omni 1.1 Flash doesn’t fix that by making the clip longer, it fixes it by making clips stackable.

The headline is “40 seconds”, but the real technical change sits somewhere quieter: how far back the model can look while it extends. There’s also a new rung on the price ladder, the 360p draft tier.

Let’s go through it. 👇🏻

What Is Gemini Omni 1.1 Flash?

A multimodal model that takes text, images and video as input and produces video and text as output. Google doesn’t frame it as a separate “video model” but as the member of the Gemini family with a video head: the same model that understands a scene also generates it.

That 57,920 output cap isn’t arbitrary. Google bills video at 5,792 tokens per second, so 57,920 tokens is exactly 10 seconds of 720p. The token limit on the model card and the clip length are two ways of writing the same number.

Where the 40 Seconds Comes From

You don’t generate 40 seconds in one call. You hand the 10-second clip you already have back to the model, say “continue”, and it generates 10 more from where it left off. You can repeat that up to 40 seconds total.

The actual improvement is in the quality of that seam. Earlier models referenced only the final second when continuing, which is why camera motion and character identity drifted at the join. Omni 1.1 reads up to 10 seconds of the prior video.

One honest caveat: Google published no comparative measurement of how well consistency holds across the full 40 seconds. Whether your character’s face after four extensions is the face from frame one is something you’ll have to test yourself.

In the API, extending is just a matter of passing the previous generation’s ID:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-omni-1.1-flash",
    previous_interaction_id=previous_video_interaction.id,
    input=[
        {"type": "text", "text": "Continue the scene."}
    ],
    response_format={
        "resolution": "360p",
    },
)

The same feature shipped in the Gemini app: every Google AI Plus, Pro and Ultra subscriber can extend a scene without writing code.

First and Last Frame Control

Pin the opening and closing frame of a shot and let the model generate the motion between them. In a plain text-to-video loop you have to describe camera movement in words; here you fix two points and leave the path to the model.

Three jobs where this pays off:

• Orbits and dolly moves. A camera circling a character 360 degrees is among the hardest shots to describe in a prompt.
• Transitions. Make one scene’s last frame the next scene’s first and join them with no cut.
• Looping clips. If the first and last frame are identical, the video loops cleanly. Directly useful for background video and web animation.

The 360p Draft Tier and the Cost Logic

This is the most practical change in the release. Generating at 360p is up to 60% faster and costs one third of 720p.

The reasoning is simple: in video generation, most first attempts are throwaway. The prompt didn’t land, the camera went the wrong way, the character changed. There is no point learning any of that at 720p.

Google’s suggested loop has three steps:

1. Draft at 360p. Generate 3-4 variations of the same scene, vary one thing at a time, compare side by side.
2. Render the pick at 720p. This is the real generation.
3. Upscale to 1080p or 4K if you need it. Delivery only.

In numbers: four variations of a 10-second scene cost $1.20 at 360p. The same four generated straight at 720p would have cost $4.

4K Is Really an Upscale

The word “4K” in the announcement can mislead. The model does not generate 4K natively; the documentation explicitly labels the output as upscaled. The default resolution is 720p and everything above it is an enlargement.

That doesn’t make 4K useless. Modern upscalers are good, and plenty of jobs demand 4K as a delivery format. Just don’t expect the extra detail native generation would bring: the $0.30 per second buys resolution, not detail.

Video References, and What Happens to Audio

You can attach up to 3 seconds of video reference to a prompt, with a maximum of 3 videos per prompt. The point is character and style consistency: “make this character perform that dance”.

Three limits to keep in mind:

• The model does not cross-reference or reason across multiple videos. You can’t hand it three references and ask what differs between them.
• Audio in a reference video is ignored.
• The model does not accept audio as input and does not edit existing audio.

On the output side there is audio generation: speech, music and sound effects. So it can create sound but cannot read it.

One more practical limit: the only language the official documentation describes as fully supported is English. Other languages have not been evaluated, so don’t assume prompts, and especially spoken dialogue, land equally well elsewhere.

Pricing

The video bill can be read two ways. The easy one is per second:

In tokens: $1.50 per million in, $9 per million for text out, $17.50 per million for video out. At 5,792 tokens per second, one second of 720p is $0.101, so both tables say the same thing.

For comparison, the fast variant of Veo 3.1 charged $0.10 at 720p, $0.12 at 1080p and $0.30 at 4K. Omni 1.1 is not a price cut at standard resolution; the saving comes entirely from the 360p rung and from throwing away fewer expensive attempts.

You can put your own workload next to Gemini, Claude and GPT prices in our LLM cost calculator, and measure how many tokens a text prompt costs by pasting it into our token counter.

What the API Doesn’t Do

The “not supported” rows on the model card matter if you were planning to treat this like a chat model:

On the Enterprise Agent Platform there is also no pay-as-you-go and no batch inference; the model runs on fixed quota and its launch stage is Preview. Google’s preview terms do allow production use, but that is not the same as the guarantees of a GA model.

No system instructions means your style rules have to be repeated inside every prompt. No function calling means this model is an agent’s tool, not the agent itself.

Where You Can Use It

• Google AI Studio: fastest way to try it with an API key.
• Gemini Enterprise Agent Platform: for enterprises, via the Agent Platform API, in the global region.
• Google Flow: open to all Google AI Plus, Pro and Ultra subscribers.
• Gemini app: scene extension is available at the same subscription tiers.

Third parties moved fast too: Adobe integrated it into Firefly and Figma into Weave, while Runway and GMI Cloud offer it on their own platforms.

Who Is It For?

• Teams building video tools. First/last frame control and extension are exactly the kind of features that sit behind a button. “Generate the transition” is now one API call.
• Anyone who iterates a lot. The 360p tier is the real win: cost per idea drops to a third.
• Ads and social content. 9:16 support plus 40 seconds of cumulative length is enough to tell a vertical story.
• Long-form video. Not this model. Forty seconds is the ceiling and consistency past four extensions is unmeasured.
• Audio-driven workflows. Also not this model. No audio input, no audio editing, and reference audio is discarded.
• Agent architectures. With no function calling and no system instructions, this can’t be the brain of an agent loop. Call it as a generation step.

Frequently Asked Questions

Q: When was Gemini Omni 1.1 Flash released? A: 27 August 2026. It went live the same day in Google AI Studio, the Gemini Enterprise Agent Platform, Google Flow and the Gemini app.

Q: How much does Gemini Omni 1.1 Flash cost? A: Per second, $0.03 at 360p, $0.10 at 720p, $0.15 at 1080p and $0.30 at 4K. In tokens, $1.50 per million input, $17.50 per million video output and $9 per million text output.

Q: How long a video can it generate? A: Ten seconds per generation. You can extend the same video in 10-second steps up to 40 seconds total.

Q: Is the 4K output real 4K? A: No, it is upscaled. The model generates 720p and the higher tiers are enlargements of it, which the documentation labels explicitly.

Q: How is it different from Veo 3.1? A: Omni 1.1 replaces Veo 3.1. The differences are scene extension that reads up to 10 seconds of context, first/last frame control, 3-second video references and the 360p draft tier. Standard 720p pricing is unchanged.

Q: Can it generate audio? A: On output, yes: speech, music and sound effects. On input, no: it does not accept audio files, does not edit existing audio, and ignores the audio in reference videos.

Q: Can I use system instructions with it? A: No. System instructions, function calling, structured output and context caching are all unsupported on this model, so persistent style rules have to live inside each prompt.

Q: Do I need an API key to use it? A: Not necessarily. Google AI Plus, Pro and Ultra subscribers can use it through Google Flow and the Gemini app within their subscription quota. The API is the paid, per-second route.

Q: What resolutions and aspect ratios are supported? A: 360p, 720p, 1080p and 4K, in 16:9 or 9:16.

Q: Is generated video labelled as AI? A: Yes, the model supports C2PA Content Credentials, so the output carries a cryptographic record that it was AI generated.

Take care… 🙂