Gemini Omni 1.1 Flash
Gemini Omni 1.1 Flash is the generative video model Google announced on 27 August 2026 and ships through Google AI Studio and the Gemini Enterprise Agent Platform. This page puts the publisher's own announcement and model card next to the settings this site's service returns for the same model, and names every point where the two do not line up.
What Gemini Omni 1.1 Flash is, in the publisher's own words
Google's announcement calls it "a production-ready update for developers that offers improved control over generative video" and lists what the update adds: extending scenes, specifying start and end frames, and generating high-resolution 4K output. The DeepMind model card, which states in its own body that it "describes Gemini Omni Flash and Gemini Omni 1.1 Flash", describes the family as "our next step towards models that can create and edit anything from any input—starting with video", with inputs of "text strings (e.g., a question, a prompt), images, audio, and video files" and outputs of "high-quality, high-resolution video with audio".
The model card lists three known limitations in the publisher's own words: "maintaining complete consistency throughout edits, generating scenes with complex motion, or rendering perfectly accurate text remains a challenge". It names SynthID, Google's digital watermarking tool, as a post-training mitigation, and states that although the model can change people's speech as part of editing, "we are restricting this capability" for now.
The card was initially published in May 2026 and last updated in August 2026, which is why the family name and the 1.1 version share one document. Nothing on this page is a statement by Google, and Google has not reviewed this page.
Quotations are left in the publisher's English. A translated specification is the usual route by which numbers start to drift.
Five modes declared for this product, four the workspace can express
This is the most consequential thing on the page, and it differs from every other video page in this series. The service this site sells through declares five task types for Gemini Omni 1.1 Flash: text to video, image to video, reference to video, video edit and video extend. The workspace's own mode vocabulary covers four task types, and two of the five the service declares have no representation in it at all.
This is a statement about what the workspace's own code can express, read from its mode list and from a search of the source for those two names. It is not a claim that editing is unavailable on the model: the publisher documents editing through conversation as a core capability, and the service sells two modes for it.
| Mode declared by the service | What it takes | Reachable in this workspace |
| Text to video | a written scene | yes |
| Image to video | a prompt and one to two stills | yes |
| Reference to video | up to ten images and up to three reference clips | yes |
| Video edit | a prompt and exactly one video | no mode of this name exists in the interface |
| Video extend | a prompt and exactly one video | no mode of this name exists in the interface |
The controls the workspace renders, and why the local table never reaches them
The workspace builds its controls from the form schema the service publishes for each mode, and only falls back to the model tables shipped in this site's source when the service publishes no such field. For this product the service publishes a resolution field, an aspect ratio field and a duration field in all three of the modes the workspace can show, so every value in the local table is dead code as far as the interface is concerned.
The middle column is what a visitor sees. The right-hand column is kept visible only so the difference is on the record; a published field wins over a local default.
| Setting | Values published by the service | Values in this site's local table, which do not reach the form |
| resolution | 360p, 720p, 1080p, 4k - default 720p | 720p, 1080p - default 720p |
| aspect_ratio | auto, 16:9, 9:16 - default auto | 16:9, 9:16, 1:1, 4:3, 3:4 - default 16:9 |
| duration | a number from 3 to 10, default 5 | 2 to 15 in steps of 1, default 5 |
| image-to-video images | 1 to 2 | 1 |
Duration is a number here, not a menu
On this model the service publishes duration as a numeric field bounded from 3 to 10 with a default of 5, rather than as a list of discrete choices. The workspace renders a number input from that, so there is no menu of durations to read off, and the local table's 2 to 15 range has nothing to do with what can be submitted.
Read against the publisher, the published bounds are coherent: the announcement describes extension arriving in ten-second increments, which implies a single generation of about that length, and the service's ten-second ceiling sits at that line. The publisher does not state a per-generation duration parameter in either reachable source, so this page does not claim to quote one.
What the publisher does state is what happens after generation. Scene extension lets the model "analyze up to 10 seconds of prior context — a leap from previous models that only referenced the final second", extending "in 10-second increments up to a total cumulative length of 40 seconds".
Scene extension runs to forty seconds, and the workspace has no extension control
Extension is the centre of the announcement and it is documented in concrete terms: ten seconds of prior context analysed, ten-second increments, forty seconds cumulative. The service sells a video-extend mode for this product with a field for exactly one reference clip, which is the shape an extension call has. Because the workspace has no mode that can express video extend, a visitor cannot reach it from these forms.
The model card adds a ceiling that is easy to miss in coverage of the launch: changing people's speech is described as a capability that is currently restricted. That is a statement about what the model is permitted to do, not about what this site sells, and it is reproduced here because edits built on speech are the kind of thing a production team plans around.
One price detail belongs in this section rather than the next. The service publishes video extend at zero points per call, labelled in its own data as a fallback. Read literally that is free, which is unlikely to be the intent. This page quotes no price of zero for it, and the question is with the operator.
Resolution: four tiers, priced by the second on both sides
Google publishes per-second prices for four output tiers, and sells a 360p draft mode that it says renders "up to 60% faster and at a third of the cost compared to Omni 1.1's standard 720p resolution". Those figures ship inside the announcement as a table image rather than as selectable text, so they were read off the image, which is archived alongside this page. The service this site sells through publishes the same four tiers.
The two ladders are not the same shape. Take the cheapest tier as 1 and the publisher's steps run 1, 3.33, 5 and 10; the service's run 1, 2.6, 4 and 7.8. That is arithmetic on four published numbers from each side and nothing more. It does not say what a point costs, and this site charges in points, quoted by the server before a task starts rather than published as a per-second rate.
The service also publishes a surcharge of 2 points per call for other output tokens, estimated at 1,000, and describes the billing mode as parameters plus duration. What that means for a given clip is a question for the quote, not for this table.
The publisher's figures are Google's own API prices in US dollars. They are not this site's prices and are shown here with attribution.
| Tier | Publisher's price, per second | This site's service, points per second |
| 360p | $0.03 | 5 |
| 720p | $0.10 | 13 |
| 1080p | $0.15 | 20 |
| 4k | $0.30 | 39 |
First and last frames, and the picker that is not offered
The announcement introduces first-and-last-frame control as a headline capability: supply the opening and closing images and "Omni 1.1 generates continuous video between two keyframes, making it ideal for complex camera orbits, zoom transitions, or seamless looping clips". The demonstration the publisher published for it, from which this page shows a frame, is the clearest example of the feature anywhere in the launch material.
The service does not declare a start-and-end mode for this product, and the local family table for it switches that mode off. A workspace that can express a start-and-end task therefore has nothing to show for this model, which is the mirror image of the previous page in this series: there, the mode was declared by the service and filtered out of another picker, and here the mode exists in the interface and the model never reaches it.
Reference clips are the one input this page has to describe carefully. The announcement says to "reference up to three seconds of video". The service publishes a field that accepts up to three reference clips. Three seconds and three files are not the same measure, and neither source contradicts the other, so both are reported as written rather than reconciled.
Sound arrives by default, and here the service says so itself
The model card describes the output as video with audio, without qualification, so audio is not an optional extra on this model. The service this site sells through goes further than it needed to and states the position in its own data for this product: sound is supported, it is generated, it is on by default, and there is no toggle. That last field is the one that decides what a visitor can do.
This is worth comparing with the previous page in this series, where the absence of an audio field had to be inferred from the fact that no such field was published. Here the service declares the absence. The visitor-facing outcome is the same in both cases, and only one of them had to be deduced.
The publisher also documents reference to voice material as an input type, and the model card lists audio among the training modalities. Neither of those puts a voice picker in this workspace, because no such field is published for this product.
What the gallery shows, and what it does not
Unlike the previous video page in this series, the publisher's footage for this model downloads: nine files were retrieved directly from Google's own storage with real payloads, including seven capability demonstrations and a partner reel. Twenty-four frames were extracted from them and every one was inspected.
Seven of those frames turned out to be the same near-white closing card carrying the model's mark, which is how all eight clips end. Five frames are the ones worth publishing: two clean generated stills from the 4K demonstration, one from the video-reference demonstration, one panel pair from the first-and-last-frame demonstration, and one frame carrying the publisher's own burned-in caption, which the caption on this page discloses.
One asset was held back for a reason that deserves naming. The file whose name announces a watermark is a partner demonstration with a platform's wordmark burned into the picture and again onto a black end card. A third-party mark inside this site's gallery would read as this site shipping watermarked output, so it is not shown and the reason is on the record.
A second asset is held back because it does not do its job. The clip published to demonstrate the 360p draft mode is itself 1920 by 1080 at the highest bitrate in the set, so the delivered file cannot show the resolution difference it exists to demonstrate. The 360p claim on this page comes from the announcement's wording instead.
Rejections and their measurements are written to this page's build record rather than left implicit in a short gallery.
Questions
Is this the official page for Gemini Omni 1.1 Flash?
No. It is a page on this site about the model. Google's announcement and the DeepMind model card are linked as sources, and neither Google nor DeepMind has reviewed this page.
Can I edit a video here, or extend one?
Not through the forms this page describes. The service declares video edit and video extend for this product, but the workspace's mode list covers four task types and neither of those two is among them.
How many modes can I actually run?
Three: text to video, image to video and reference to video. The service declares those three plus the two editing modes the interface cannot express.
Why does the duration field read three to ten seconds when the publisher talks about forty-second scenes?
They measure different things. Three to ten is the range the service publishes for a single generation; forty seconds is the cumulative total the publisher documents for scene extension, reached in ten-second increments. This workspace has no extension control.
Is 4K available?
The service publishes four output tiers for this product, 360p, 720p, 1080p and 4k, with 720p as the default. The publisher prices all four, at $0.03, $0.10, $0.15 and $0.30 per second. What this site charges is quoted in points before a task starts and is not a per-second rate.
Do the videos have sound?
Yes, by default, and there is no switch. The service states in its own data for this product that audio is supported, generated and on by default, with no toggle. The publisher's model card likewise describes the output as video with audio.
How many reference images or clips can I attach?
The service publishes up to ten images and up to three reference clips for reference to video, and one to two images for image to video. The publisher separately documents referencing up to three seconds of video, which is a duration and not a count.
Why does the gallery have five items when the publisher shipped nine files?
Because eight of the nine end on the same branding card, and most of the remaining frames carry interface chrome or burned-in captions. Two files are not represented at all: one carries a partner platform's visible watermark, and the other is the 360p demonstration, which is published at 1920 by 1080 and so cannot demonstrate the resolution it exists to show.