Google shipped a video-generation update on August 27, 2026 that changes how much control developers get over AI-made clips. Gemini Omni 1.1 Flash now extends scenes to a cumulative 40 seconds, lets creators pin an exact first and last frame, and upscales draft footage all the way to 4K. The update landed through Google’s developer blog and the official Gemini API changelog, and it was picked up within hours by MarkTechPost, which called the shift a move from “a capable generator” to “a directable one.”
The headline number is context: Omni 1.1 can now read up to 10 seconds of a prior clip when extending it, compared with roughly one second in earlier Omni releases. Google’s own promotional material draws a direct line against Veo, its more cinema-oriented model line, which it says only carries one second of context into an extension. That single spec is doing most of the marketing work in this release, and it is also the spec worth scrutinizing, because “40 seconds” is not one continuous render. It is a chain of 10-second segments stitched together, each one drawing on the tail of the last.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Gemini Omni 1.1 Flash actually changed
Three features anchor this release, according to Google’s developer blog post and the Gemini API release notes. First, scene extension now uses a 10-second context window instead of referencing only the final frame of the previous clip, which Google says produces better visual consistency across cuts. Second, first-and-last-frame control lets a user specify both the opening shot and the closing shot, and the model generates the motion path between them. That is the feature enabling camera orbits, dolly-zooms, and loop animations that previously required trial-and-error prompting. Third, output resolution now spans 360p, 720p, 1080p, and 4K, with 360p positioned as a fast, cheap draft tier and the higher resolutions produced through upscaling rather than native high-resolution generation.
The base generation ceiling has not moved: a single render is still capped at 3 to 10 seconds, per the Gemini API documentation, which also lists a 24 frames-per-second output rate and a context window of 1,048,576 tokens for the underlying model. What changed is the chaining mechanism around that ceiling. Google’s recommended workflow, described in its own materials and echoed by outlets covering the release, is a three-step pipeline: draft at 360p, pick the best take and render it at 720p, then upscale the final selection to 1080p or 4K only at the delivery stage. That sequencing matters commercially, because 360p drafts are described as roughly one-third the cost of a standard 720p generation and up to 60% faster, which changes the economics of iterating on a shot before committing to a final render.
Why the 10-second context window is the real story
Most AI video tools generate short clips well but struggle to maintain continuity when a user asks for a sequel shot. A character’s shirt changes color, a background prop disappears, camera framing drifts. That inconsistency is the practical wall between “cool demo” and “usable production tool,” and it is the wall Google is targeting with the expanded context window. Reading 10 seconds of prior footage instead of a single final frame gives the model far more information about motion trajectory, lighting continuity, and subject identity when it generates the next segment.
Google AI’s official account framed the comparison bluntly in its launch post: “With Omni 1.1 you can extend scenes based on 10 seconds of context from your original video, a big jump from Veo’s 1 second!” (Google AI, X). That framing is notable because Veo is Google’s own higher-end cinematic model line — the company is explicitly positioning its faster, cheaper Omni Flash tier as having a technical edge over its own flagship in this one dimension, even as Veo retains advantages elsewhere, including shorter per-clip cost tuning through Veo 3.1 Lite, which offers configurable 4, 6, or 8-second clips with cost scaling to duration.
First and last frame control: what it unlocks
Keyframe-style control has existed in some form across the AI video field for a while, but Google’s implementation ties it directly into the same pipeline as scene extension. A user sets a starting frame and an ending frame, and Omni 1.1 fills in the motion between them. Google’s own account described the mechanic on X: “Set your starting shot and ending frame, and Omni generates the continuous motion in between” (Google, X). In practice, this is the feature that lets a creator storyboard a shot the way a filmmaker or animator would, specifying compositional endpoints instead of describing camera movement in a text prompt and hoping the model interprets it correctly.
The model also accepts up to three seconds of video as a style or motion reference, which developers can use to keep a character’s appearance or a scene’s visual tone consistent across separate generation calls. Combined with what Google calls stateful multi-turn editing — the ability to iteratively adjust an existing clip’s motion or style rather than regenerating it from scratch — the toolset starts to resemble a lightweight, API-driven editing suite rather than a one-shot generator. That is a meaningful repositioning for a “Flash” tier model, which is typically marketed as the fast, cheap option rather than the full-featured one.
Resolution, upscaling, and the honest limits of “4K”
It is worth being precise about what “4K” means in this release, because the marketing framing and the technical documentation say slightly different things depending on the source. The Gemini API’s official model page lists output video at 360p, 720p, 1080p, and 4K, 24 frames per second. Separately, Google’s release notes for the Gemini API state plainly that 1080p and 4K outputs are generated using upscaling, with 720p functioning as the base, “true” render resolution. A standalone Gemini Omni Flash upscaler tool referenced in coverage of the release allows uploads of MP4, MOV, or MKV files up to 50MB and 30 seconds long, with a choice of 2x or 4x enlargement and the credit cost shown before processing begins.
That distinction, native generation versus upscaled output, is not unique to Google. Most AI video systems currently generate at lower base resolutions and rely on separate upscaling passes for anything above 1080p, since native high-resolution video diffusion remains computationally expensive at scale. The practical implication for developers is that 4K output from Omni 1.1 Flash should be evaluated as an enhanced-detail final pass rather than a from-scratch 4K render, which matters for anyone benchmarking fine detail, text legibility, or fast-motion sharpness against models that claim native high-resolution generation.
Pricing and the economics of iterative video generation
Google has not published an itemized dollar-per-second or dollar-per-clip rate card for Omni 1.1 Flash alongside this update, at least not in any material publicly available at the time of writing. What is documented is relative cost. The 360p draft tier is described in coverage of the release as roughly one-third the cost of standard 720p generation, positioned specifically so that developers can iterate on composition, motion, and framing cheaply before committing credits to a full-resolution render. Upscaling to 1080p or 4K is treated as a separate, additional metered step, with credit cost displayed to the user before the upscale runs.
That tiered-cost structure is a deliberate response to a problem every AI video product has faced since the category emerged: generation is expensive per attempt, and creative work is inherently iterative. A model that only offers one price point for one resolution punishes experimentation. By pricing 360p drafts at roughly a third of the 720p cost and treating high-resolution output as an optional final step, Google is effectively encouraging a workflow where most of the trial and error happens cheaply and the expensive step is deferred until a creator has already settled on a take.
Competitive comparison: Omni 1.1 Flash vs. Veo, Sora-class models, and Kling
The AI video generation field has split into two rough camps over the past year: models chasing longer, native continuous shots and cinematic fidelity, and models chasing controllability, editability, and API-friendly iteration. Google’s own Veo line, alongside Sora-class systems and Kling, has generally been described in coverage as leaning toward the cinematic, longer-single-pass camp, with an emphasis on visual fidelity per render. Omni 1.1 Flash, by contrast, is being positioned as the directable, developer-oriented option: chainable segments, explicit frame-level control, and a tiered resolution pipeline built around iteration rather than a single expensive render.
The trade-off is real. A single Omni 1.1 Flash generation still tops out at 10 seconds, and the 40-second figure is a cumulative chain rather than one continuous pass, meaning visual drift across a full 40-second sequence is a realistic risk even with the expanded 10-second context window. Models built around longer native single-shot generation avoid that stitching risk entirely, at the cost of less granular control over individual segments. Which approach wins out for a given use case depends heavily on whether the priority is uninterrupted visual continuity or precise directorial control over composition and pacing.
| Model / Line | Max scene extension | Context used per extension | Native max resolution | Frame control |
|---|---|---|---|---|
| Gemini Omni 1.1 Flash | 40 seconds (chained) | 10 seconds of prior footage | 720p native, 4K via upscale | First/last frame control |
| Google Veo 3 (flagship line) | Extension supported, shorter continuity per Google’s own comparison | ~1 second of prior footage | Higher native fidelity per clip | Prompt-driven, less granular |
| Veo 3.1 Lite | Single clip, cost-tuned | Not applicable (single pass) | Configurable, cost scales with duration | Prompt-driven |
| Sora-class / Kling-style models | Vendor-dependent, often longer single-pass | Vendor-dependent | Native higher-resolution output claimed | Varies by vendor, generally less API-exposed |
The table above reflects Google’s own public comparisons of Omni 1.1 Flash against its Veo line, plus general positioning described in coverage of the release; exact specs for third-party models such as Sora-class systems and Kling were not part of Google’s verified release material and should be checked against each vendor’s own documentation before being used for procurement decisions.
Technical specifications at a glance
For developers evaluating whether to integrate Omni 1.1 Flash into a production pipeline, the concrete limits documented in Google’s Gemini API materials are the numbers that matter most, since marketing framing around “40 seconds” and “4K” can obscure the underlying mechanics.
| Spec | Value | Source |
|---|---|---|
| Single generation duration | 3–10 seconds | Gemini API video docs |
| Cumulative extension length | Up to 40 seconds (chained, 10-second increments) | Google developer blog |
| Prior-context window for extension | Up to 10 seconds | Google developer blog |
| Frame rate | 24 FPS | Gemini API model page |
| Resolutions supported | 360p, 720p, 1080p, 4K | Gemini API model page |
| Native vs. upscaled resolution | 720p native; 1080p and 4K via upscaling | Gemini API changelog |
| Video reference input | Up to 3 seconds, for style/motion guidance | Coverage of Google developer blog |
| Model context window (text/tokens) | 1,048,576 tokens | Gemini API model page |
| Upscaler tool file limits | Up to 50MB, 30 seconds, MP4/MOV/MKV | Gemini Omni Flash upscaler tool |
Historical context: how we got from static-frame extension to 10-second context
AI video generation only became commercially usable in the past two to three years, and scene extension specifically has been one of the field’s persistent weak points. Early text-to-video systems could produce a strong single clip but broke down the moment a user asked for continuity into a second clip, because the model had no memory of the first beyond perhaps a single reference frame. That is precisely the limitation Google is calling out in its own prior Omni releases, which it says referenced only the final second of a clip when extending it. The jump to a 10-second context window is not a new capability category so much as a substantial widening of an existing one, and it reflects the broader trend across the AI video field toward longer effective memory during generation, mirroring similar context-window expansions that reshaped text-based large language models over the past several years.
The tiered draft-to-final resolution pipeline also echoes a pattern already established in AI image generation, where low-resolution draft passes followed by targeted upscaling became standard practice once generation costs at scale became a real constraint for high-volume users. Applying that same pattern to video, where per-second compute costs are considerably higher than for a single image, is a logical next step, and it suggests other major video model vendors are likely to adopt similar tiered-cost structures if they have not already.
Market impact: what this means for creators, studios, and developers
For independent creators and small production teams, the combination of cheaper 360p drafting and first/last-frame control lowers the cost of iterating toward a usable shot, which is often the single biggest barrier to using AI video tools in a real production rather than a demo reel. Being able to storyboard with explicit start and end frames, rather than relying purely on text prompting to control camera movement, brings the workflow closer to how traditional animation and previsualization already operate, which should shorten the learning curve for people coming from established creative tools.
For developers building on top of the Gemini API, the update matters less for any single feature and more for the direction it signals: Google is treating its Flash-tier model, not its flagship Veo line, as the primary surface for iterative, developer-facing video tooling. That is a notable allocation of engineering effort, and it suggests Google expects the bulk of API-driven, high-volume video generation demand to come through cost-sensitive, iteration-heavy workflows rather than one-off cinematic renders. For competitors, the message is that context window and chaining mechanics, not just raw output fidelity, are becoming a differentiator worth competing on directly.
What developers should evaluate before adopting Omni 1.1 Flash
Teams considering a switch to or trial of Omni 1.1 Flash should test a handful of specific things before committing production workflows to it. First, visual drift across a full 40-second chained sequence should be evaluated directly, since the 10-second context window reduces but does not eliminate the risk of gradual inconsistency across multiple stitched segments. Second, the actual cost delta between 360p drafting and 720p or upscaled final renders should be measured against a team’s own typical iteration count, since Google’s one-third-cost figure for 360p is a general claim rather than a guaranteed outcome for every prompt type. Third, teams that need native high-resolution output, rather than upscaled 4K, should confirm whether the upscaling pipeline meets their fidelity bar for the specific use case, particularly for text-heavy frames, fine texture, or fast motion where upscaling artifacts are more visible.
It is also worth testing the first/last-frame control against a range of motion types, since Google’s own materials highlight camera orbits, dolly-zooms, and loop animations as the flagship use cases, which suggests the feature may be tuned most heavily for those specific motion patterns rather than arbitrary complex scene transitions.
Predictions: where AI video generation goes from here
- Expect competing video model vendors to publish their own context-window comparisons within the next few months, using the same “seconds of prior context” framing Google introduced against its own Veo line, since it has proven to be an easily understood marketing metric.
- Tiered draft-to-final resolution pricing, already standard in AI image generation, is likely to become the default cost structure across most major AI video products within the next year, as vendors look for ways to make iterative workflows economically viable at scale.
- Native high-resolution generation, as opposed to upscaled output, will likely remain a premium, higher-cost tier across the industry for the near term, since the computational cost of native 4K video diffusion has not shown signs of falling as quickly as image generation costs did.
- First/last-frame and keyframe-style control is likely to become a baseline expectation for professional and semi-professional AI video tools, following the same trajectory as inpainting and outpainting did in AI image generation.
- Chained scene extension with longer context windows will keep pushing toward the practical limits of consistency, but full native long-form generation without stitching artifacts is unlikely to be commercially available at consumer-friendly cost within the next year, given current compute constraints.
How Omni 1.1 Flash fits into Google’s broader Gemini strategy
This release did not happen in isolation. Google has spent much of 2026 pushing Gemini deeper into both consumer surfaces and developer tooling, and video generation has become one of the more visible battlegrounds for that push. Positioning Omni Flash, rather than the higher-end Veo line, as the model getting the most aggressive new controllability features suggests Google is optimizing for developer adoption and API call volume over marquee cinematic showcases. That is consistent with the company’s broader pattern of using Flash-tier models as the default entry point for third-party integrations, where cost and iteration speed tend to matter more than peak output fidelity.
It also puts pressure on the rest of the field to match not just output quality but workflow ergonomics. A model that is slightly behind on raw visual fidelity but meaningfully ahead on cost-efficient iteration and precise control can still win developer mindshare, particularly for teams building consumer-facing products where the priority is a fast, cheap draft-to-final pipeline rather than the single most impressive demo clip.
Frequently asked questions
What is Gemini Omni 1.1 Flash?
It is Google’s updated multimodal video generation and editing model, released August 27, 2026, adding a 10-second context window for scene extension, first/last frame control, and 4K upscaling on top of the existing Gemini Omni Flash line.
How long can a video be with Omni 1.1 Flash?
A single generation is 3 to 10 seconds. Using scene extension in 10-second increments, a chained sequence can reach a cumulative 40 seconds, though this is a stitched chain rather than one continuous render.
Is the 4K output from Omni 1.1 Flash native 4K?
No. Per Google’s Gemini API release notes, 1080p and 4K outputs are generated using upscaling from a 720p base render, not native high-resolution generation.
How does Omni 1.1 Flash compare to Veo?
Google’s own materials say Omni 1.1 reads up to 10 seconds of prior context when extending a scene, compared with roughly one second for Veo. Veo generally remains positioned toward higher per-clip cinematic fidelity, while Omni Flash targets iterative, cost-tiered, developer-controllable workflows.
What is first/last frame control?
It lets a user specify the exact starting frame and ending frame of a clip, and the model generates the motion in between, enabling more deterministic camera moves such as orbits, dolly-zooms, and seamless loops.
How much does Omni 1.1 Flash cost?
Google has not published an exact public rate card alongside this update. What is documented is relative pricing: 360p draft generation is roughly one-third the cost of standard 720p generation, and upscaling to 1080p or 4K is billed as a separate, additional step.
Where can developers access Omni 1.1 Flash?
The model’s video capabilities are documented in the Gemini API, alongside a dedicated Gemini Omni Flash upscaler tool for enlarging existing footage up to 4K.
What is video reference input used for?
Omni 1.1 Flash accepts up to 3 seconds of video as a style or motion reference, which developers can use to keep a character’s appearance or a scene’s tone consistent across separate generation calls.
