What is Gemini Omni — "The Next Step Beyond Veo" as Shown at I/O 2026
Let's start with the big picture. Gemini Omni is a video generation and editing model that Google CEO Sundar Pichai and Google DeepMind showcased as a core topic at the keynote of Google I/O 2026, which opened on May 19, 2026. Google's official announcement describes it in a single line: "create anything from any input — starting with video." The first version made generally available was the lightweight, high-speed "Gemini Omni Flash," which launched globally on the same day.
What matters here is that Omni is not simply "a new version of a video generation tool." Until now, Google's generative media was divided into separate lanes by function — Veo for video, Imagen for images, a different pipeline for audio, and so on. Omni folds all of these into a single model, integrating the "intelligence (reasoning and world knowledge)" of Gemini itself with the "rendering power" of the media models. Nicole Brichtova, Director of Product Management at DeepMind, told TechCrunch that this represents "the next step in the progress of combining Gemini's intelligence with the rendering capabilities of our media models." The official blog post was authored by Koray Kavukcuoglu, CTO of DeepMind and Chief AI Architect at Google.
A concrete example makes this clearer. In the demo presented by Kavukcuoglu, simply giving the instruction "explain protein folding as a clay animation" generated a single stop-motion video complete with accurate narration audio. With a single photo on hand, you can use it as a starting point to create a video, or edit the photo using text — an experience reminiscent of Google's image editing model "Nano Banana." In other words, Omni behaves like a collaborator that "takes whatever you put in, thinks it through, and returns a finished piece of footage."
Pichai framed this direction as a historic inflection point in AI. In his words, "through world models, AI is moving from predicting text to simulating reality." This single sentence is the backbone for understanding Omni. Below, we dive into five key topics that every creator should know.
Bidirectional & Simultaneous Multimodal Reasoning — Thinking About "Everything You Paste" All at Once
The technical core of Omni is that it is "natively multimodal." Rather than sorting different types of data — text, images, audio, and video — into separate steps and then stitching them together, a single core neural network reasons across all of them simultaneously within the same forward pass (a single inference run). With the conventional relay approach of "passing text model output to a media model," context was lost at modality boundaries and artifacts (breakdowns) tended to appear at the seams. Omni eliminates those boundaries altogether.
For creators, the practical benefit ties directly to "freedom with reference materials." In Google's words, "Omni transforms any reference — image, text, video, or audio — into a single cohesive output." A still image for a character's appearance, a separate video clip for nuances of movement, an audio sample for mood, and text for instructions — all of these can be mixed together and fed into a single prompt. The model reasons across all of them and returns a single video that reflects every element. This is what "bidirectional and simultaneous" actually means in practice. Not only are inputs multimodal, but outputs are also headed toward multimodality in the future (discussed later) — making it a true any-to-any system in the fullest sense.
However, as of the current time (early June 2026), audio input is officially noted as starting with "voice reference" only, with other audio input types to be rolled out gradually. This is worth noting without overstating.
Creator Tips: Every official and media prompt test converges on the same golden rule: "attach reference materials whenever possible." A text-only prompt forces the model to invent visual identity from scratch, and the more editing rounds you go through, the more randomness accumulates. Conversely, providing even one reference image, motion clip, or audio track dramatically improves output stability. If you want to lock in a character, the emerging best practice is to first create a "character sheet" image using Nano Banana (the image model) and then reuse it as a reference across all scenes. Design a character once, then summon them into any scene — this "design then summon" mindset is becoming the foundation of character management in the Omni era.
Physical Engine Intelligence — How "World Models" Are Transforming the Common Sense of Video
The biggest reason Omni is called a generational leap rather than "an extension of Veo" lies in its understanding of the laws of physics. Google's official description states that Omni possesses "an improved intuitive understanding of forces such as gravity, kinetic energy, and fluid dynamics," and "combines intuitive understanding of physics with Gemini's knowledge of historical, scientific, and cultural context." In his keynote, DeepMind CEO Demis Hassabis introduced Omni as a "world model" — a system that builds an internal understanding of reality and reasons about what should happen next within a given scene.
Why does this work? Conventional video generation has primarily relied on pattern-matching vast numbers of pixels to predict "the next frame." Even if the result looks plausible, the behavior is inconsistent. Characters morph between cuts, shadows ignore light sources, and water flows like a texture rather than a physical substance — early Sora's infamous examples of fountain water flowing upward or objects passing through walls are emblematic of this. Omni, rather than guessing "the next pixel," is said to directly incorporate a physical framework of how forces operate into the generation process itself.
Concrete demos make a compelling case. The standout example highlighted across various media is a "glass marble" clip, in which a marble rolls down a complex Rube Goldberg-style track, with synchronized sound effects firing each time it bounces or a bell rings. One review described it as "physics you can believe in." Kavukcuoglu's clay animation protein explainer is another prime example of "generation backed by scientific knowledge," in terms of narrative accuracy. A demo of a professor writing a mathematically correct derivation of trigonometric functions on a blackboard has also been reported — demonstrating that the model consistently handles the mechanics of the hand, chalk pressure, and the logical ordering of steps.
Tips for Creators: The strong physics understanding means that even without detailed prompts about "how things should move," you'll get back natural falling, collisions, splashing water, and flowing hair or fabric. This reduces the burden on creators while giving a significant boost to educational and explainer content. For product videos, it's worth deliberately targeting physical depictions that used to break down — like "liquid being poured into a container and foaming up" or "a metal ball dropping onto a water surface with spreading ripples." Conversely, if you want to intentionally break real-world physics (such as cartoonish exaggerated expressions), you'll need to explicitly add style directives (e.g., "cartoon style," "ignoring gravity") to override the world model's "seriousness."
Google Flow Integration ―― Pro Editing Tools Become "Conversations"
The professional face of Omni is its integration into Google's generative video production studio, "Google Flow." At I/O 2026, Flow was upgraded with four enhancements in addition to Gemini Omni Flash support: Flow Agent, major improvements to Flow Tools and Flow Music, and a mobile app. Since this is the area where creator workflows will change the most, I want to take a careful look.
At the center is Flow Agent — a "creative assistant" built on Gemini models that, in Google's words, "plans and reasons through complex tasks based on your input and under your control." In practice, it handles dialogue suggestions, plot proposals, simultaneous generation of multiple variations, batch editing of assets, and intuitive renaming and organization of collections. It's positioned as a partner that brings "deep project understanding" to every stage from brainstorming through production and editing.
Flow Tools is a system for assembling custom workflows in natural language without writing code, with the ability to share your own tools with other users and remix each other's work. Flow Music is equally powerful — Omni now lets you direct music videos through conversation, enabling fine-grained edits such as rewriting lyrics, rebuilding specific sections, and style-converting entire tracks while preserving the melody and arrangement (style covers). Mobile apps are available for both Flow and Flow Music, supporting on-the-go production.
Flow usage is managed through "Flow Credits" tied to pricing tiers. Based on figures compiled by various media outlets, the allocation is: AI Plus at Flow 200 / Flow Music 3,000; AI Pro at 1,000 / 10,000; AI Ultra (5x) at 10,000 / 30,000; and AI Ultra (20x) at 25,000 / 30,000 (pricing details in the next section).
Creator Tips: The true power of Flow Agent lies in using it to "generate multiple options simultaneously and choose from them." Rather than refining a single cut one option at a time, generating batches of variations — different lighting, different camera angles — and then conversationally refining the best result is ultimately faster. With Flow Tools, converting your standard go-to processes (such as cropping to vertical 9:16 with branded-color captions) into a tool once means your team or community can reuse it, making it highly effective for high-volume production work. Flow Music's "style conversion while preserving melody" pairs well with marketing use cases where the same track needs to be tailored for different target audiences.
Live Streaming Edit — A New Editing Loop That Sculpts Video Through Conversation
The greatest experiential impact Omni has had on creators is that "video editing has become as easy as having a conversation." Google itself titled Omni's introduction page "Create and edit videos like you're having a conversation." This is what this article calls "live streaming edit" — an editing loop in which you sculpt footage in real time, going back and forth in dialogue.
Traditional generative video was a "gacha (slot machine)" approach: you threw a prompt and regenerated all clips from scratch. With Omni, you can fix just part of a scene using natural language instructions. The official prompt guide explains: "You can ask Omni for specific updates only — like changing the background or adding a new caption — without needing to re-prompt the entire scene," and "retain the video across multiple rounds of revisions, keeping what's working." Each turn's instructions stack on top of the previous turn, allowing edits to proceed while maintaining consistency in characters, lighting, and objects. One review described it as "feeling like you're talking to an intelligent collaborator rather than operating a sophisticated slot machine" — a reference to precisely this back-and-forth feel.
A real-world example introduced by CineD illustrates this well. Simply saying "when the character touches the mirror, make the mirror ripple beautifully like liquid" rewrites only that single moment while preserving character continuity and scene logic. The sense of "fixing footage through conversation" rather than "reshooting it" is beginning to reshape the assumptions behind editing.
That said, some measured caveats are in order. Character consistency across multiple rounds of editing has historically been a weakness in this category, and CineD cautions that it should be "verified before relying on it for production work." Additionally, if an editing prompt is vague, unintended areas may also change — a pitfall already familiar to Nano Banana users, and one that TechCrunch likewise flags.
Creator Tips: The golden rule for editing instructions is "specific, one thing at a time." Instead of "make it better," name the target and intent explicitly — for example, "add backlight coming from the window at the far left, emphasizing the character's outline." Cinematic vocabulary works well for camera movement — the official guide recommends terms such as "push in," "dolly zoom," "locked off," "natural smartphone zoom," and "webcam style," and provides examples of sequential camera instructions like "quick tilt up from a close-up of the shoes to a medium shot, then out to wide." When consistency starts to break down, it's faster to stop trying to talk your way through it and instead return to the last successful frame or a reference image to rebuild from there.
Project Astra ―― Toward a Resident Visual Assistant
The fifth axis is "Project Astra," a separate line from Omni itself but closely intertwined with it. This is a research prototype developed by Google DeepMind toward a "universal AI assistant," aiming for a resident assistant that understands the world captured by a camera in real time and processes conversation and vision simultaneously. Some overseas media and blogs refer to it as "Project Astra 2.0," but the official name on Google DeepMind's page is simply "Project Astra" — it is worth noting that "2.0" circulates not as an official product brand but as a colloquial term referring to advanced, next-generation capabilities. This article will use both names for convenience.
In terms of capabilities, it understands objects in context while highlighting "what to focus on right now" on screen, and responds on the spot without delay or interruption. A "proactive" behavior — initiating conversation on its own — is also a defining feature. Regarding memory, it retains recent video frames, past queries, and cross-device context within a session, recalling past conversations for personalized responses. The benchmark of "approximately 10 minutes of in-session memory" discussed since early demos has been carried forward in a more refined form. Tool integration is also implemented, enabling task completion on the user's behalf — including operating Search, Gmail, Calendar, and Maps, as well as interface control.
For deployment, Google has explicitly stated its intention to extend Project Astra's capabilities to Gemini Live, a new Search experience, and glasses as a new form factor. In fact, several of the latest features in Gemini Live were first explored in Project Astra. Eyewear brands such as Warby Parker and Gentle Monster are reported as partners for eyewear, while Samsung (Android XR) is named as a partner for XR hardware; Android XR audio glasses are slated to arrive "this fall." A specialized version is also being developed in partnership with Aira, a visual assistance service, for users with visual impairments or low vision.
Creator-perspective TIPS: Astra has the potential to transform the "entry point" of video production. Capturing reality live through the camera of glasses or a smartphone, and bridging on-the-spot subjects, locations, and motion as "reference material" to Omni — when the loop of "see → shoot → edit through conversation" connects into a single flow, the effort involved in location scouting and reference gathering is greatly reduced. At this point, Astra and Omni remain separate layers, but it is worth keeping in mind the direction of this integration emerging through Gemini Live as a starting point.
Pricing and Access — From Free YouTube to $200/month Ultra
Where and how much you can use Omni follows the Google AI subscription structure revamped at I/O 2026. As a free entry point, users aged 18 and over can try Omni Flash at no cost via YouTube Shorts' "Remix" feature and the YouTube Create app. To use it in earnest within the Gemini app and Google Flow, one of Google AI's paid plans is required.
Pricing breaks down as follows: AI Plus at $7.99/month, AI Pro at $19.99/month, and the higher-tier AI Ultra in a two-tier structure — "Ultra 5x" with five times the usage quota at $99.99/month, and "Ultra 20x" at $199.99/month. The top Ultra tier has been reduced from the previous $250/month to $200/month, and with the addition of the new $100/month five-times quota tier, the range of upper-tier options has expanded. Ultra 5x includes 20 TB of cloud storage and a YouTube Premium individual plan. Read alongside the Flow credit allocations mentioned in the previous section, a clear segmentation emerges: Plus for "those who want to try it," Pro as "the practical line for individual creators," and Ultra for "high-volume, commercial workflows."
One important note for commercial use is the watermark that is always applied to output. Every video generated by Omni has Google's invisible digital watermark "SynthID" embedded in it, which can be verified in the Gemini app, Gemini in Chrome, and Search. This cannot be opted out of, and in the API discussed later, it is expected to be "mandatory" rather than "optional," alongside C2PA Content Credentials. While this aligns with the societal need to identify AI-generated content, it is worth factoring in at the estimation stage that it may pose a constraint for certain commercial workflows that presuppose clean output.
How Silicon Valley Covered It — Positioning Against Seedance and Sora
Silicon Valley's reception has focused on "qualitative transformation of the experience" rather than "flashy features." TechCrunch emphasized the breadth of the roadmap from its headline itself — "Turning images, audio, and text into video — and this is just the beginning." The Verge introduced Omni as a new class of models aimed at "creating anything," moving away from the narrow constraints of previous video generation. VentureBeat discussed it as an "any-to-any" model, analyzing its disruptive power as an end-to-end workflow for enterprises (advertisers and production companies). CineD, targeting cinematographers, welcomed the ability to animate one's own digital avatar with one's own voice as a "production time-saver," while calmly noting that Google is deliberately holding back broader audio editing capabilities — a consideration of the risks of dialogue alteration.
Coverage of the competitive landscape is realistic, avoiding excessive hype. The shared assessment on launch day is that it is "not the highest-fidelity model available," with multiple comparative articles noting that Seedance 2.0 still holds the top spot on fidelity leaderboards and that Sora 2 remains stronger in certain physics-based scenarios. Nevertheless, Omni is valued not for competing in an image quality race, but for carving out a new arena: an editing experience of "talking with an intelligent collaborator." TechCrunch cited Luma AI (which generates ad campaigns from product briefs) as a comparable startup building agentic, multi-step creative workflows, and positioned Omni as "Google's serious move for consumers."
The tension between its "two faces" — consumer and enterprise — is also a point of discussion. While Google pitches avatars to consumers as "personalized memes" for self-creating scenes like moon trips or award ceremonies, Brichtova's emphasis on text-rendering accuracy in advertising also hints at genuine enterprise ambitions. It should also be noted that some reporting is circulating for which primary-source confirmation of the operational status of certain competing services could not be obtained; this article is limited to facts that could be independently verified.
Essential for Creators — How to Design Prompts and Build Consistency
I'd like to bundle the individual arguments so far into "patterns" that work in actual production. What Google DeepMind's official prompt guide repeatedly emphasizes is the philosophy that "you don't need to over-instruct Omni." In the guide's own words: "Tell it what you want to create, then watch the model's reasoning and world knowledge fill in the details." A good prompt, it says, "reads like a clear brief to a talented collaborator, not like a legal contract."
Building on that, the official guide lists the following axes of control when you want to be specific: shot framing and movement (wide / medium / close-up), style (realistic / cinematic, grounded / majestic), lighting (crisp / warm / ethereal), location, and action. Community-side testing has found that prompts answering four questions — what you're making / what inputs to use / what to keep consistent / what the final video is for — produce stable results, and it has been shared internally at Google that "users who covered six dimensions got dramatically better outputs." This is less a suggestion and more practical knowledge that separates merely "using the model" from truly "mastering the model."
As I have repeated throughout this article, the way to achieve consistency comes down to one thing: "attach references, and design your character before you summon them." Whether the material is real-world footage or something created in Nano Banana, providing a single reference image lets you reuse it across scenes. When using an avatar, there is a dedicated onboarding flow for deepfake prevention that asks users to record a short self-video reading a series of numbers — it's best to view this extra step as a safety design that guarantees "proof of identity" for commercial use. In the final editing stage, following just three rules makes a significant difference in yield during mass production: be specific and change one thing at a time, describe the camera in cinematographic terms, and when something breaks, roll back to the most recent successful frame.
What's Coming Next — API, Omni Pro, Image/Audio Output & Glasses
Finally, let us organize the developments visible as of early June 2026 in order of likelihood. The most imminent is API availability for developers and enterprises: Google has announced delivery "within weeks," and various outlets expect rollout to begin in mid-to-late June. The path will be two-pronged — the Gemini API for individual developers and Vertex AI for enterprises. Reports indicate that the API at launch will support text/image/audio/video input with video output, multi-turn conversational editing, and AI avatars, with SynthID and C2PA Content Credentials mandatorily embedded in all outputs.
In the medium term, expansion of output modalities has been pledged. Omni starts "with video first," but Google has explicitly stated it will extend over time to image, text, and even audio output — TechCrunch relayed a future vision of "generating images from audio and audio from video." Longer clip lengths (beyond the current 10-second ceiling) and higher resolution are also said to be in development. One point worth noting precisely and without exaggeration: the 10-second limit has been explicitly described not as an architectural constraint but as a product decision made "to get it into more people's hands faster."
Further down the road stands the higher-tier model Gemini Omni Pro. It is set to launch when it delivers what can be called a "step change" over Flash, with no specific timeline given. Various outlets consider it likely to debut first in a new AI Ultra tier at $100/month, accompanied by longer clips and higher resolution. In parallel, the Gemini 3.5 line — which serves as Omni's intelligence layer — continues to evolve. Gemini 3.5 Flash, introduced at I/O, has become the default model for the app and AI Mode, while the higher-tier Gemini 3.5 Pro has been announced for rollout the following month (June 2026). On the form-factor front, the aforementioned Android XR audio glasses are slated to arrive "this fall," making it the next focal point to watch how Project Astra's persistent vision and Omni's generation and editing capabilities will be bridged together.
Overall, the next milestones Silicon Valley creators should watch for come down to four: (1) the explosion of tool integrations triggered by API availability in late June; (2) the moment any-to-any comes closer to completion with the expansion into image and audio output; (3) the lifting of length and resolution constraints with Omni Pro; and (4) whether the fall glasses rollout makes the "shoot-and-edit-instantly" loop a reality. Omni was not the image-quality king on day one. But that, in itself, reflects a strategy in which Google moved first to claim the ground of "how you interact with video" — rather than competing on image quality.