Google Genie 4: The AI World Model That Talks Back

Direct answer: Google DeepMind has unveiled Genie 4, a research preview of an interactive world model that generates playable environments with native sound and lets you reshape them mid-session just by describing what you want. Scenes now stay coherent for around two minutes, physics and human characters look more believable than in Genie 3, and the same technology is already being tested for robot training in simulation. It is not a product you can buy off the shelf today, but it is a clear sign of where interactive AI is heading — and why small businesses should start thinking about simulation, not just video.

What happened

On 27 October 2026, Google DeepMind announced Genie 4 as a research preview, the successor to Genie 3, which had arrived in the same research-preview form in August 2025. The announcement was covered the same day by TechCrunch and Ars Technica, and the framing in both reports is consistent: this is not a finished consumer tool, and it is not a simple upgrade to a video model. It is a step towards AI systems that do not just describe or depict a world, but sustain one while you interact with it.

The headline change is audio. Earlier Genie versions were largely silent worlds. Genie 4 generates sound natively alongside the visuals, covering three kinds of audio that matter in different ways. There is music, which sets mood and pace. There is ambient sound, the background texture that makes a place feel occupied — wind, traffic, a workshop hum, rain on a roof. And there is interactive dialogue, meaning characters inside the world can speak and respond rather than simply moving silently through a scene. For anyone who has tried a silent simulation, the difference is immediate. Sound is what turns a moving picture into a place.

The second headline change is real-time steering in natural language. While a Genie 4 world is running, you can type or say something like change the weather, move the camera, or add a person or an object, and the world adjusts without restarting. That sounds like a small convenience, but it changes the workflow completely. In older generative video, if you wanted a different sky or a different angle, you usually regenerated the whole clip and hoped the rest stayed the same. Here, the world is meant to persist while one element changes. That persistence is the whole point of a world model, and it is also the hardest part to get right.

The third change is stamina. Genie 4 scenes remain coherent for up to around two minutes, roughly double the roughly one minute Genie 3 could manage. Two minutes still sounds short if you are thinking in terms of films or training courses, and it is. But in world-model terms, doubling coherent duration is significant, because every extra second is another chance for the model to forget where something was, drift away from the original prompt, or let physics quietly break. Longer coherence means longer walkthroughs, longer practice scenarios, and longer demonstrations before the illusion needs a reset.

Under the surface, DeepMind also reports better physics, including aerodynamics and collisions, and more natural human characters. That matters because the previous generation of generated people often gave the game away: stiff walks, odd hands, faces that seemed to slide rather than express. More natural characters do not just look nicer. They make dialogue believable, they make training scenarios usable, and they make a prototype feel like something a customer could actually react to rather than laugh at.

Two supporting technologies are worth naming because they explain how Genie 4 tries to stay consistent. Gemini Omni is used to anchor appearance and behaviour, so a character or object is more likely to keep looking and acting like itself as the scene continues. Gemini 3.8 Flash acts as an in-loop guiding agent, helping steer the generation while the world is running. You do not need to remember those names to use the idea, but they explain why DeepMind is talking about controllability, not just prettiness. The model is being guided while it runs, not simply left to improvise.

There is also a robotics thread that is easy to miss but may prove the most important part of the announcement. DeepMind described Gemini 4 Argon, a robotics experiment in which the technology was used with Apptronik’s Apollo humanoid robot playing checkers in simulation. That is not a claim that a robot is about to walk into your warehouse and start working. It is a demonstration of the underlying logic: if you can build a believable, interactive simulated world, you can let a robot practise inside it before it ever touches the real world. Simulation is cheaper than breakage, safer than trial and error on live equipment, and repeatable in a way real life rarely is. For businesses that train people or operate physical equipment, that logic will feel familiar.

Access, for now, is limited and staged. Genie 4 is initially available to trusted testers, with wider access planned through Google AI Ultra via Project Genie, and through the Gemini API and AI Studio for early-access partners. There is also a waitlist. That staged rollout is normal for a research preview, and it is worth stating plainly: most UK small businesses will not be logging into Genie 4 this week, and anyone promising immediate access or a guaranteed price is getting ahead of the official position. The sensible response is to watch, learn the vocabulary, and test the adjacent tools that are already available, rather than building a business plan around a preview.

All Genie 4 outputs carry a SynthID watermark, Google’s marker for AI-generated content. That is a detail with practical consequences. Watermarking helps platforms and viewers identify synthetic material, and it limits how cleanly generated scenes can be passed off as real footage. For a small business, that is not a reason to dismiss the technology, but it is a reason to be honest about what is simulated and what is real, especially in marketing or training where trust matters.

Coverage of the launch also places Genie 4 in a fast-moving field. Runway’s GWM Worlds 2, World Labs working with AMD, reported talks involving Decart and Amazon, Rivian’s work with Third Dimension, and Google’s own Veo video work are all part of the same broader race: who can build believable, controllable, interactive worlds first, and at a cost others can afford. Genie 4 does not end that race. It raises the bar for what counts as a world, rather than a clip.

For background on how AI video tools have already reached small businesses, see our coverage of the Kling 4.0 API launch and our practical guide to AI video marketing for UK small businesses. Genie 4 sits one step beyond those tools: less finished, more interactive, and potentially more useful for testing ideas before you spend money filming them.

How Genie 4 actually works

It helps to start with a plain definition, because the term world model gets used loosely. A world model is an AI system that learns, from vast amounts of video and related data, how environments tend to look and behave, and then uses that learned understanding to generate a new environment that responds to input. It is not a game engine with hand-built levels. There is no artist placing every chair and scripting every door. The model infers what a street, a workshop, a shop floor or a garden should look like, how objects should move, and what should happen when you act.

In Genie 4, the process typically starts with a prompt, an image, or both. You might describe a small café on a rainy evening, or supply a photo of a product display and ask for a world built around it. The model generates the environment in real time, meaning it is rendering the next moments as you move through it, not playing back a pre-rendered file. You can navigate the space, look around, and interact, and the model keeps generating what should plausibly come next.

That real-time generation is why audio matters so much here. In a pre-rendered video, sound can be added afterwards by an editor. In an interactive world, sound has to be generated in step with what is happening, or the illusion breaks. If a door opens, the sound should arrive with the movement. If a character speaks, the words should match the situation and the timing. If the weather changes because you asked it to, the soundscape should change too. Native audio means the model is generating sound as part of the world, not bolting it on at the end. That is technically harder, and it is why DeepMind is presenting it as a headline feature rather than a footnote.

Real-time steering works in a similar way. When you ask for a weather change, a camera move, or a new person or object, you are not editing a video timeline. You are giving the guiding system a new instruction while the world continues. Gemini 3.8 Flash, acting as the in-loop guiding agent, helps interpret and apply that instruction without collapsing the rest of the scene. Gemini Omni, meanwhile, helps keep the identities in the scene stable: the same character should still look like the same character after the rain starts, and the same product should still look like the same product after the camera moves. Anyone who has used generative video will know how often that stability fails. A logo morphs, a face changes, a chair quietly becomes a different chair. Anchoring is DeepMind’s answer to that drift, and the two-minute coherence claim is, in effect, a claim that the anchoring now holds for longer.

Physics is the third pillar, and it is the least glamorous but often the most revealing. Better handling of aerodynamics and collisions means objects fall, bounce, slide and collide in ways that look plausible rather than cartoonish. A ball should not pass through a table. A door should not swing through a wall. A vehicle should not float. These details sound trivial until you try to use a simulation for training or prototyping. If physics is wrong, people learn the wrong thing, or customers notice the fakery before they notice the product. Improved physics does not make Genie 4 a physics simulator in the engineering sense, but it makes the worlds usable for a wider range of rough-and-ready tests.

It is worth distinguishing Genie 4 from Veo, Google’s video generation family, because the two are often confused in headlines. Veo is built to produce high-quality video clips, often with cinematic control over camera, lighting and motion. You prompt it, it renders a clip, you watch it. Genie 4 is built to produce an environment you inhabit and alter. You prompt it, it builds a world, you move through it and change it. Veo is closer to filmmaking. Genie is closer to a rehearsal room, a showroom, or a training ground. Both have value, but they solve different problems. A business that needs a polished thirty-second advert will still reach for a video tool first. A business that needs to test ten versions of a shop layout, or let a trainee practise a tricky conversation, will find the world-model idea more interesting.

The robotics experiment helps clarify that distinction. In the Gemini 4 Argon demonstration, Apptronik’s Apollo robot plays checkers in simulation. Checkers is deliberately simple: a board, pieces, rules, turns. The point is not the game. The point is that a robot can practise perception, planning and interaction inside a generated world where mistakes cost nothing, and where the scenario can be reset or varied instantly. Scale that idea up, and you get the long-term promise: robots and AI assistants that learn in simulation before they operate in the real world, and human workers who can practise alongside them in the same simulated space.

A timeline helps put Genie 4 in context. Genie 3 arrived as a research preview in August 2025, offering interactive worlds without native audio and with roughly a minute of coherence. Over the following year, rivals pushed their own world-model announcements, including Runway’s GWM Worlds 2 and World Labs’ work with AMD, while reports of Decart and Amazon talks and Rivian’s collaboration with Third Dimension showed that carmakers, chipmakers and cloud providers all see simulation as strategic, not decorative. Genie 4, in October 2026, answers with sound, steering, longer coherence and a robotics demonstration. The next steps, based on the access plan DeepMind has described, are trusted-tester feedback, then wider availability through Google AI Ultra via Project Genie and through the Gemini API and AI Studio for early-access partners. No one should read that as a promise of a specific date for general availability, because DeepMind has not given one in this preview announcement.

The comparison table below summarises the shift in plain terms. It is based only on what DeepMind and the launch coverage have stated, not on hands-on testing, since this is a research preview with limited access.

That table also shows what has not changed. Genie 4 is still a research preview. It is still short-form. It still watermarks its outputs. It still has the limits discussed below. The progress is real, but it is progress within a research frame, not a declaration that interactive worlds are solved.

One more distinction matters for anyone evaluating the news critically. A world model can look impressive in a curated demo and still be fragile in everyday use. Demos tend to show the best prompts, the best camera paths, and the most forgiving scenarios. Real use involves awkward prompts, unusual products, crowded scenes, and users who do things the demo never tried. The two-minute figure is an upper bound for coherence, not a guarantee that every scene will last two minutes without drift. That is why trusted-tester feedback matters, and why small businesses should treat the announcement as a signal to explore, not as a purchase order.

Sources for this section include the launch reporting from TechCrunch (Google’s Genie 4 world model adds audio, more realistic worlds and real-time steering) and Ars Technica (DeepMind’s Genie 4 world model adds audio and more realistic people to simulations), alongside DeepMind’s own preview description as reported there.

What it means for UK small businesses

For most UK small businesses, Genie 4 will not change anything this week. That sentence is worth saying first, because AI news often implies an urgency that the access position does not support. If you run a café in Leeds, a plumbing firm in Bristol, a boutique in Manchester or an accountancy practice in Glasgow, you cannot rebuild your marketing around a tool that is still in trusted-tester preview. What you can do is understand where this technology points, and start using the cheaper, available tools that already capture part of the value, so that when world models do become broadly accessible, you are not starting from zero.

The most immediate idea is prototyping. Small businesses constantly face small, expensive decisions: how should a new shop layout feel, where should a product display go, what should a refurbished waiting area look like, how would a market stall read from across the street? Today, answering those questions often means sketches, mood boards, or paying someone to mock something up. A world model, even in rough form, lets you walk through a version of the idea, change the lighting or the weather, move a counter, add a person, and see whether the idea survives contact with a spatial reality. You would not use that walkthrough as a final architectural plan, but you would use it to kill bad ideas early, before you spend money on fixtures, signage or a refit. For a business where a refit costs thousands, killing one bad idea pays for a lot of experimentation.

A concrete walkthrough helps. Imagine a small garden centre outside York planning a new indoor plant section. Today, the owner might look at a floor plan and a few supplier photos. With a world-model approach, the owner could describe the section — low wooden benches, hanging plants, a gravel path, soft daylight — then walk through it, ask for rain outside the windows, add a customer browsing, and check whether the path feels cramped or the signage is readable from the entrance. If the path feels cramped in simulation, it will feel cramped in real life. If the owner then tries a wider path and a lower bench, the comparison is immediate. None of that replaces a real site visit or a builder’s advice, but it shortens the gap between an idea in someone’s head and a shared picture the whole team can react to.

Product demos are the second area. Many small businesses sell things that are hard to show in a flat photo: a piece of furniture in a room, a kitchen layout, a piece of equipment in use, a holiday let, a training course. A playable world lets a potential customer explore the context around the product, not just the product itself. A furniture maker could let a buyer see a chair in different rooms and lighting. A kitchen fitter could let a homeowner walk a proposed layout and open cupboards. A holiday-let owner could let a guest explore the cottage and garden before booking. The current two-minute limit means these demos would need to be short and focused — a single room, a single route, a single decision — rather than a sprawling virtual tour. But short and focused is often exactly what a buyer needs. A two-minute walkthrough that answers one question well is more persuasive than a ten-minute video that answers none.

Training is the third area, and it may be the most practical in the medium term. Small businesses train staff in customer service, safety, handling complaints, and operating equipment, usually by shadowing, role-play, or a slide deck. A world model offers a middle ground: a repeatable scenario where a trainee can practise a difficult conversation, a safety walkthrough, or a till-side upsell, and where the scenario can be varied — a calm customer, then a rushed one, then a confused one — without needing three different volunteers. The interactive dialogue in Genie 4 is relevant here. A trainee who can actually talk to a simulated customer, rather than clicking through multiple-choice options, learns timing, tone and recovery in a way a quiz cannot teach. The limits still apply: dialogue accuracy is imperfect, and a simulation must never be treated as a substitute for real supervision or certified safety training. But as a rehearsal tool before the real thing, it has obvious appeal, especially for businesses that cannot afford bespoke e-learning development.

Marketing video is the fourth area, but it needs the most caution. It is tempting to imagine Genie 4 replacing video shoots: describe a scene, generate a world, record a walkthrough, post it. For rough social content, behind-the-scenes teasers, or concept videos, that workflow may eventually work. For a polished brand advert, it does not yet, for three reasons. First, two minutes of coherence is not the same as two minutes of broadcast-ready footage; drift and audio errors will show. Second, SynthID watermarking means outputs are marked as AI-generated, which is honest but limits how they can be presented as real footage. Third, customers can tell when a place is simulated, and for local businesses, trust is built on real premises, real staff and real products. The smart marketing use is therefore not replacement but augmentation: use simulation to test a concept, storyboard an idea, or create a clearly labelled teaser, then film the real thing once the concept is proven. Our guide to AI video marketing for UK small businesses makes the same point for current video tools: AI works best when it supports a real story, not when it pretends to be one.

Costs and access limits deserve a frank paragraph, because this is where small businesses most often get caught out. There is no public price for Genie 4 in this preview announcement, and no UK general-availability date. Access is staged through trusted testers, then Google AI Ultra via Project Genie, and through the Gemini API and AI Studio for early-access partners, with a waitlist for others. Google AI Ultra is a premium tier in Google’s existing structure, but the preview announcement does not set out a Genie-specific price, and this article does not invent one. Any third party offering guaranteed Genie 4 access for a fee should be treated with scepticism until Google confirms the route. The practical cost question for a small business is therefore not what Genie 4 costs today, but what the surrounding workflow will cost when it arrives: compute time, staff time to prompt and review, and the time to check outputs for errors before they reach customers. Those costs are real even when a tool looks free, and they are why a short, focused use case beats a vague plan to use AI for everything.

There is also a skills point that small businesses should not underestimate. Prompting a world model is a different skill from writing a social media caption. It involves describing spaces, lighting, behaviour and camera movement clearly, and then steering the result without breaking it. The businesses that benefit first will not necessarily be the ones with the biggest budgets. They will be the ones that build a small habit: one person who experiments, documents what works, and shares short examples with the team. That habit can start now with currently available video and image tools, including the Kling API work covered in our Kling 4.0 API launch article. When world models become accessible, that person will already know how to brief a scene, spot drift, and judge whether a simulation is good enough to show a customer.

Sector examples make the abstract concrete, with the usual caveat that these are possibilities, not promises. A trades firm could simulate a boiler-room layout to check access for servicing before quoting. An estate agent could offer a short, clearly labelled simulated walkthrough of a proposed staging, alongside real photos, not instead of them. A nursery or tuition centre could rehearse a parent-tour route. A small manufacturer could test how a new workstation layout feels before moving heavy equipment. A restaurant could test table spacing and a new menu-board position. In each case, the value is not photorealism. It is decision support: seeing a spatial idea before committing money to it.

None of this means a small business should wait passively. The useful preparation is boring and cheap. Keep a list of decisions that would benefit from a walkthrough. Save good photos of your premises, products and displays, because future world tools will work better from real reference images than from vague descriptions. Practise writing short, specific scene briefs: what room, what light, what people, what action, what camera path. And keep a clear line between simulated and real content in anything customer-facing. That discipline — label simulation as simulation — will matter more as these tools improve, not less.

Finally, there is a competitive point. Larger rivals will experiment with world models first, because they have staff to spare and budgets for premium tiers. Small businesses cannot outspend them, but they can out-focus them. A large chain might build a generic virtual store. A small business can build a precise simulation of its actual shop, its actual products, and its actual customer journey, and use it to make one better decision. That focus is the advantage small businesses have always had: knowing their customers and their premises intimately. World models, when they arrive, will reward that knowledge rather than replace it.

Honest limits / what to watch

Enthusiasm about Genie 4 needs to be balanced with what DeepMind and the launch coverage actually flag as limits, because those limits decide whether a given use case is viable or merely exciting.

Drift is first. Over a longer session, the world can slowly wander away from its starting point. A room changes shape, a product changes colour, a character’s behaviour becomes less consistent. The two-minute figure is a measure of how long coherence can hold, not a promise that it always will. For a short, single-purpose walkthrough, drift may be tolerable. For a longer training scenario or a customer-facing demo, drift can quietly teach the wrong thing or misrepresent a product. The practical response is to keep sessions short, reset often, and check the output against reality before using it.

Object permanence is closely related. In a real room, an object stays where you left it unless someone moves it. In a generated world, an object can vanish, duplicate, or change when the camera moves away and back. That matters for any use case that depends on inventory, layout or safety: a fire exit that disappears, a tool that moves, a display that changes. Until permanence improves, simulations should be treated as sketches, not as records.

Legible text inside the world is another known weakness. Signs, labels, menus, price tags and instructions often render with garbled or unreadable lettering. That is a serious limit for retail, hospitality and training uses, where text carries meaning. A simulated shop with unreadable signage does not test signage. A simulated manual with garbled instructions does not train anyone. For now, text-heavy scenes should be avoided, or text should be added afterwards by a human in a clearly labelled overlay, not relied upon from the model itself.

Audio accuracy is the fourth limit, and it is easy to overlook because the presence of sound feels like progress on its own. Generated dialogue can mispronounce, mistime, or misunderstand context. Ambient sound can be plausible but wrong: the wrong kind of traffic, the wrong workshop noise, music that does not fit the brand. For training dialogue, wrong tone can teach the wrong habit. For marketing, wrong sound can undermine trust faster than a visual error. Audio should be listened to critically, not just noticed.

There are also broader questions to watch, beyond the technical limits. One is cost and access. Research previews often become premium products first, and small businesses may face a gap between announcement and affordable access. Another is intellectual property and likeness: generating worlds that include recognisable products, brands or people raises questions that a small business should not improvise on. A third is disclosure: as SynthID watermarking shows, platforms and regulators increasingly expect AI-generated material to be identifiable, and customer trust will favour businesses that label simulation honestly. A fourth is data: what reference images or business details you feed into a world tool, and where they go, will matter once these tools are used with real premises and real products.

What to watch over the next few months is therefore specific, not vague. Watch for trusted-tester reports that go beyond curated demos: how often does drift appear, how usable is dialogue in practice, how well does a real shop or product survive as a reference? Watch for the access rollout: when does Project Genie reach Ultra subscribers more broadly, and what do API and AI Studio partners actually build with it? Watch for rival responses, especially from Runway, World Labs and others in the world-model race, because competition tends to drive down cost and improve reliability faster than any single announcement. And watch for the first credible small-business case studies, not from Google’s marketing, but from businesses similar to yours describing a decision they made better because they simulated it first.

The bottom line is measured. Genie 4 is a genuine advance in interactive AI: sound, steering, longer coherence, better physics, more natural people, and a robotics link that shows why simulation matters beyond entertainment. It is also a research preview with known limits, limited access, watermarked outputs, and no public price or UK availability date. Small businesses should learn from it, prepare for it, and test the adjacent tools that are already available — without building plans, budgets or customer promises on a preview. The businesses that benefit most will be the ones that treat Genie 4 as a rehearsal room for ideas, not as a replacement for real premises, real staff, and real products.

FAQs

What is Google Genie 4?

Genie 4 is Google DeepMind’s latest interactive world model, released as a research preview on 27 October 2026. It generates a playable simulated environment from a prompt or image, complete with native audio, and lets you steer it in real time using natural language. Unlike a fixed video, the world continues to generate as you move through it and change it.

How is Genie 4 different from Genie 3?

The biggest differences are sound, steering and stamina. Genie 4 adds native audio including music, ambient sound and interactive dialogue, supports real-time language steering such as changing the weather or camera or adding a person or object, and holds a scene together for around two minutes — roughly double Genie 3’s one minute — with improved physics and more natural-looking people.

Can UK small businesses use Genie 4 today?

Not generally, no. At launch it is available to trusted testers first, with wider access planned through Google AI Ultra via Project Genie and through the Gemini API and AI Studio for early-access partners. There is a waitlist, but no public price or UK general-availability date has been announced for this preview, so businesses should watch the rollout rather than assume immediate access.

Is Genie 4 a video generator like Veo?

Not quite. A video generator produces a fixed clip you watch. A world model produces an environment you can move through and change as you go. Genie 4 is closer to a playable simulation than to a finished film, which is why its current value is mainly in prototyping, training and exploration rather than polished final video. Veo remains the closer comparison for finished, cinematic clips.

What are Genie 4’s main limits?

DeepMind flags drift over longer sessions, imperfect object permanence, difficulty with legible text inside the world, and audio that is not always accurate. Outputs also carry a SynthID watermark. Those limits mean simulations should be kept short, checked against reality, and clearly labelled as simulated, especially for customer-facing or safety-related uses.

CTA: Get a Free AI Consultation