Omni 1.1 Flash: Google's Powerful, Essential 4K Upgrade

Omni 1.1 Flash is Google’s production release of the generative video model it launched in preview earlier this year, and it landed on 27 August 2026 with the feature developers had been asking for since June: output above 720p. The model now writes 360p, 720p, 1080p and 4K, extends an existing clip to forty seconds, and interpolates a shot between a start frame and an end frame you supply.

Google DeepMind product managers Anish Nangia and Alisa Fortin announced it as a “production-ready update for developers that offers improved control over generative video”. That phrasing is doing real work. The previous release was a capable demo with one resolution and one editing verb. This one is the version you would actually wire into a product, and Google has priced it accordingly — paid tier only, no free quota, billed by output token rather than by clip.

The resolution headline deserves a closer read than it usually gets, and this article gives it one. Google’s own API reference lists 1080p and 4K as upscaled outputs, not native renders, which changes what you should expect from a 4K frame and what you should budget for it. That is not a criticism so much as a specification, and it is the kind of detail that decides whether a model fits a client deliverable or does not.

What follows covers exactly what shipped in Omni 1.1 Flash, how the pricing arithmetic works when the meter runs on tokens, what the forty-second extension ceiling really means, the surfaces the model is live on today, and the list of things it still refuses to do — including a regional restriction that lands squarely on anyone working in the UK or the EU. If you build video tooling, sections three and six are the ones that will change your plans.

Table of contents

What Google Shipped in Omni 1.1 Flash
The 4K Headline in Omni 1.1 Is an Upscale, Not a Native Render
How Omni 1.1 Flash Pricing Actually Works
Scene Extension Is the Most Substantial Change in Omni 1.1
Where You Can Use Omni 1.1 Flash Today
What Omni 1.1 Flash Still Cannot Do
What Partners Are Building on Omni 1.1 Flash
What Omni 1.1 Flash Changes for Teams Planning Video Work
References

What Google Shipped in Omni 1.1 Flash

Five changes went into this release. They are not equally significant, and the one Google led with is not the one that will alter most workflows.

Scene extension that reads ten seconds back

Extension takes an existing clip and continues generating from where it stopped. The change in Omni 1.1 is how much of the clip the model reads before it does that. Earlier models referenced only the final second; this release analyses up to ten seconds of prior context. Google’s stated result is improved visual consistency and narrative adherence across the join.

First and last frame interpolation

You can now supply the starting frame and the ending frame of a shot and have the model generate the footage between them. Google positions this for camera orbits, zoom transitions and seamless loops — the shots where the endpoints are known and the middle is the hard part.

A 360p draft tier

The new bottom rung is the practical one. A 360p render comes back up to 60% faster — Google footnotes that as system throughput of 360p against 720p — and costs a third of what the standard 720p tier costs. It exists for storyboard iteration and rapid prototyping, not for delivery.

1080p and 4K output

The top of the ladder is the release’s headline. Omni 1.1 Flash accepts resolution values of 360p, 720p, 1080p and 4k inside the response_format object, with 720p as the default. Google’s marketing language is “upscale up to 4K resolution”, and the API reference is more precise still, which we will come back to.

Video references in the prompt

The model already accepted text and image references. Omni 1.1 adds video: up to three clips of up to three seconds each, used to carry motion and character likeness into a new scene. Any audio inside a video reference is ignored, and the model does not reason across multiple videos at once.

Capability	Omni 1.1 Flash	Stated limit
Prior context read when extending	Up to 10 seconds	Previous models read the final second only
Total clip length after extension	40 seconds cumulative	Appended in 10-second increments
Native generated clip	3 to 10 seconds at 24fps	16:9 default, 9:16 supported
Output resolutions	360p, 720p, 1080p, 4K	720p is the default value
Video references per request	Maximum 3 clips	Up to 3 seconds each, audio ignored
Uploaded video for edit or extend	10 seconds or less	Longer clips only via multi-turn
Context window	1,048,576 tokens	Text, image and video input

The 4K Headline in Omni 1.1 Is an Upscale, Not a Native Render

This is the part of the release most coverage flattens, and it is worth being exact about because it determines what a 4K frame from this model will and will not contain.

What the documentation actually says

Google’s Gemini API reference lists the four accepted resolution values with a one-line description each. Two of those descriptions carry a parenthetical: 1080p is documented as “1080p output (upscaled)” and 4k as “4K output (upscaled)”. The model card lists output video as “3s-10s (360p/720p/1080p/4K, 24 FPS)” without separating the two groups, but the API reference is unambiguous.

Why the distinction matters for production

An upscale cannot add detail the generation stage never produced. It gives you a delivery-ready frame size, a cleaner edge on text and geometry, and a file that drops into a 4K timeline without a resize — all of which are genuinely useful. What it does not give you is more information in the image. If your shot depends on fine texture holding up under scrutiny, the ceiling is set at the render stage, not at the export stage.

Where the native resolution still shows

Google’s own guidance leaks the boundary. The documentation recommends delivery="uri" rather than inline base64 “for videos larger than 4MB (>720p when available)” — an aside that treats 720p as the reference point around which everything else is sized. The 360p tier is described in throughput terms against 720p as well. Throughout the docs, 720p is the model’s centre of gravity, and 1080p and 4K sit on top of it.

What this changes about how you brief it

Practically, it means you prompt for a shot that survives an upscale. Wide geometry, strong silhouettes and clean motion scale well. Dense small-scale texture and legible on-screen detail do not — and the model card independently flags text rendering accuracy as a known weakness, which an upscale will enlarge rather than repair.

How Omni 1.1 Flash Pricing Actually Works

The meter on this model does not run the way video pricing usually runs, and misreading it is an easy way to build a business case on the wrong number.

Billing is by token, not by second

Omni 1.1 Flash is billed on total output token consumption. Google’s pricing page lists $17.50 per million output video tokens and $9.00 per million output text tokens on the paid tier, with input at $1.50 per million tokens across text, image, video and audio. The per-second figure everyone quotes is derived, not charged: Google states a rate of 5,792 tokens per second of 720p video, which works out at approximately $0.10 per second.

There is no free tier

Both the generally available gemini-omni-1.1-flash and the gemini-omni-flash-preview model list “Not available” against every free-tier price row. If you are prototyping, the 360p tier is your budget control, not a free quota. Provisioned throughput is not supported either, so committed-capacity pricing is off the table for now.

The 360p tier is the real cost lever

At a third of the 720p cost, a 360p draft runs at roughly $0.033 per second by Google’s own stated ratio. Generating four variations at 360p before committing one to 4K is the workflow the release is built around, and Google’s example applications — a draft room that generates three or four variants varying one thing at a time — make that explicit.

Model and tier	Stated price per second	Resolution basis
Omni 1.1 Flash, 360p draft	About $0.033 (one third of 720p)	Derived from the stated ratio
Omni 1.1 Flash, standard	About $0.10	5,792 tokens per second of 720p
Veo 3.1 Lite	$0.05 / $0.08	720p / 1080p, no 4K output
Veo 3.1 Fast	$0.10 / $0.12 / $0.30	720p / 1080p / 4K
Veo 3.1 Standard	$0.40 / $0.60	720p and 1080p / 4K

Omni 1.1 against Google’s own Veo line

Veo 3.1 publishes explicit per-resolution rates; Omni 1.1 publishes one token rate anchored to 720p. That means the two are not cleanly comparable above 720p, and anyone modelling a 4K workload should treat the Omni figure as a floor rather than a quote. On the 720p line where both are documented, Omni 1.1 Flash matches Veo 3.1 Fast exactly and sits at a quarter of Veo 3.1 Standard.

Published price per second of generated video, scaled to the $0.60 top rate

Veo 3.1 Standard, 4K $0.60 — 100% of the top rate

Veo 3.1 Standard, 720p and 1080p $0.40 — 67%

Veo 3.1 Fast, 4K $0.30 — 50%

Omni 1.1 Flash standard, and Veo 3.1 Fast at 720p $0.10 — 17%

Veo 3.1 Lite, 720p $0.05 — 8%

Omni 1.1 Flash draft tier about $0.033 — 6%

Bars are each published rate divided by the $0.60 Veo 3.1 Standard 4K rate. The draft-tier figure is one third of the $0.10 standard rate, as stated by Google.

Scene Extension Is the Most Substantial Change in Omni 1.1

Resolution got the headline. Extension is the change that alters what you can actually make.

Ten seconds of context instead of one

The old behaviour was structurally limiting: a model that reads only the last frame of a clip has no way to know what the shot was doing before that frame. Lighting drifts, a character’s clothing changes, a camera move reverses. Reading ten seconds gives the model the motion vector and the visual grammar of the shot it is continuing, which is why Google frames the improvement as consistency and narrative adherence rather than raw quality.

The forty-second ceiling

Extensions append in ten-second increments to a cumulative maximum of forty seconds. A native clip is three to ten seconds, so the practical shape is one generation plus three extensions. Forty seconds is not a film, but it is a complete product demo, a social cutdown, or a single scene in something longer — and it is four times what most competing single-request generations offer.

The rules that constrain extension

Extension only appends to the end of a clip. You cannot prepend footage, and you cannot extend from the middle. Uploaded videos must be ten seconds or shorter unless you are working multi-turn on a clip the model generated itself. And you cannot extend an uploaded video in which somebody is speaking to add more dialogue — the character can stay silent, or you can work multi-turn on generated footage instead.

The stateful editing loop behind it

Extension rides on the same Interactions API that powers conversational editing. You pass previous_interaction_id rather than re-uploading the clip, and the model keeps the video state between turns. One caveat is buried in the best-practice notes: setting store=false for faster synchronous generation means the result cannot be edited in a later turn. Speed and editability are a trade, and you choose per request.

Seconds of video, scaled to the 40-second cumulative ceiling

Maximum cumulative clip length 40 seconds — 100% of the ceiling

Length added per extension turn 10 seconds — 25%

Prior context the model now reads 10 seconds — 25%

Longest native single generation 10 seconds — 25%

Each video reference clip 3 seconds — 8%

Prior context earlier models read 1 second — 3%

Every bar is a figure Google publishes, divided by the stated 40-second cumulative maximum.

Where You Can Use Omni 1.1 Flash Today

The rollout is broader than the previous release and split across developer and consumer surfaces.

Google AI Studio and the Gemini API

The primary developer entry point is Google AI Studio, calling gemini-omni-1.1-flash through the Interactions API with the Gen AI SDK for Python or JavaScript, or raw REST. The convenience field interaction.output_video is SDK-only; a direct REST caller pulls the base64 video out of the steps array instead.

Gemini Enterprise Agent Platform

Enterprises build against the same model through the Agent Platform API, which is where Google’s named customers are already running production workloads. The model is generally available on the paid tier there, with the older preview identifier still listed alongside it.

Flow and the Gemini app

On the consumer side, Omni 1.1 is available to Google AI Plus, Pro and Ultra subscribers globally in Google Flow, starting the day of the announcement. Scene extension specifically is available to the same subscriber tiers in the Gemini app. If you have wondered what the difference between Google’s AI subscription tiers buys you, this release is a clean example of the answer.

Surface	Who it is for	What you get
Google AI Studio	Developers	Full Omni 1.1 model, paid tier, all resolutions
Gemini Enterprise Agent Platform	Enterprises	Same model via the Agent Platform API
Google Flow	AI Plus, Pro and Ultra subscribers	Omni 1.1, globally, from launch day
Gemini app	AI Plus, Pro and Ultra subscribers	Scene extension, globally

What Omni 1.1 Flash Still Cannot Do

Google publishes an unusually candid limitations list, and several entries will decide whether the model fits a given build.

The regional restrictions are real

Editing or extending an uploaded video is not available to users in the European Economic Area, Switzerland or the United Kingdom. Editing and extending video the model generated itself works everywhere. Uploading and editing images containing minors is likewise unavailable in those three territories, and images containing certain recognisable people are blocked outright. For a UK or EU studio wanting to run client footage through the model, that is not a footnote — it is the specification.

Voice and audio are locked down

The model generates audio alongside video from your text description, but voice editing is not supported and audio references cannot be uploaded at all. Google’s model card says speech manipulation is restricted pending further safety research. Audio inside a video reference is ignored, so a reference clip contributes motion and likeness only.

The parameters that do not exist

System instructions, temperature, top_p, stop sequences and negative prompts are all unsupported. Google’s workaround for the last one is to put negatives in the prompt as plain language — “no dialogue”, “no extra sound effects”. YouTube videos cannot be used as a media source, and multi-video reasoning is explicitly called out as producing degraded or unexpected output.

The quality limitations Google names itself

The model card lists three: difficulty maintaining consistency across edits, trouble with complex motion scenes, and imperfect text rendering. Omni 1.1 Flash is not a computer vision system bolted onto a renderer — it is a native multimodal model that reasons about physics and world knowledge — but reasoning about a scene and rendering it flawlessly are still two different problems. English is fully supported; other languages have not been evaluated.

Restriction	Scope	Workaround
Editing or extending uploaded video	Blocked in the EEA, Switzerland and the UK	Work on model-generated footage instead
Adding dialogue when extending uploads	Not supported where a person is speaking	Keep the character silent, or go multi-turn
Voice editing and audio references	Unsupported globally	Describe the audio in the text prompt
Negative prompts, temperature, top_p	Parameters do not exist	Write negatives as plain prompt language
Reasoning across multiple videos	Explicitly unsupported	One reference video per creative intent
Provisioned throughput	Not offered on this model	Standard pay-as-you-go only

What Partners Are Building on Omni 1.1 Flash

Google named four customers already running the model in production, and their comments are more informative than the usual launch-quote boilerplate.

Adobe, Figma and Runway

Adobe has integrated Omni Flash into Adobe Firefly for video editing. Figma’s creative director for Weave, Itay Schiff, framed the upgrade in directing terms: “With extensions, richer reference material, and 4K resolution, Gemini Omni Flash takes teams beyond generating videos to truly directing them.” Runway’s chief creative officer, Jamie Umpherson, described it as fitting the existing prompt-image-video entry points his users already work through.

The reliability argument from GMI Cloud

The most specific endorsement came from GMI Cloud’s VP of marketing, Louisa Guo, who pointed at accuracy rather than features: “For customers creating educational and explanatory content, where getting things right is essential, that reliability matters more than any single feature.” That is the argument for a model built on Gemini’s world knowledge rather than a pure generation stack.

What the applications suggest

Google’s own showcase applications point at three shapes: a transition generator that takes a first and last frame, a property walkthrough that moves a camera through rooms without inventing furniture, and a draft room that generates several 360p variants side by side. All three are thin products around one API capability, which is a fair signal of where the near-term commercial opportunity sits for teams building on this model.

What Omni 1.1 Flash Changes for Teams Planning Video Work

If you commission or build video tooling, four things follow from this release.

Draft cheap, finish expensive

The 360p tier changes the economics of iteration, and any workflow that ignores it pays roughly three times over on every discarded variant. Build the draft-then-promote loop into the product rather than leaving it to whoever is writing the prompts.

Model the cost before you commit

Token-based billing means your finance model needs the token rate, not the per-second convenience figure — particularly above 720p, where Google publishes no separate rate. Anyone comparing this against other providers should work from published token rates rather than headline per-clip prices, the same discipline that applies across the wider LLM API pricing landscape.

Provenance is not optional

Every clip carries a SynthID watermark, invisible to viewers and detectable programmatically. If you publish generated video, build detection into your own review process rather than relying on downstream platforms — the same argument that applies to verifying AI content with SynthID across images and text.

Treat the region rule as an architecture decision

If your users are in the UK or the EU and your product depends on editing footage they upload, Omni 1.1 Flash cannot do that job today. Design around generated footage, or route that specific capability elsewhere. This is a good moment to revisit how model choice fits your broader AI strategy rather than treating it as a procurement detail.

Where it sits in Google’s line-up

Omni 1.1 Flash is the conversational, multimodal, iterate-fast option; Veo 3.1 remains the higher-priced production renderer with explicit per-resolution pricing. The two are complementary rather than competing, and the same pattern is visible across Google’s recent model releases — including Gemini 3.5 Transcribe, which took the same specialised-model approach to audio. Our AI models and tools hub tracks the rest of the line-up, and the marketing services team can advise on where generated video belongs in a content programme.