What is Google Genie 4, and how is it different from AI video?

Think of a video generator as a printer. You give it a prompt, it hands back a finished clip, and you watch. Genie 4 works more like a live set. The scene keeps rendering while you move through it with keyboard controls, and your text prompts change it mid-run.

What Genie 4 can do that Genie 3 could not Genie 3 arrived in August 2025 as DeepMind’s first world model with real-time interaction, running at 720p and 24 frames per second according to its Wikipedia entry. Genie 4 changes four things you’d notice straight away. First, live steering. You can type a change while the world is running: bring in rain, swing the camera, drop in a new person or object. DeepMind says Gemini adapts your prompt to fit the surrounding scene, so a “stranger appears” request doesn’t paste someone awkwardly into the frame. Second, sound. This is the first Genie model that generates audio itself. Wind in trees, waves at a beach, music, all produced alongside the picture. Third, dialogue. Characters can talk to each other and react to what you do or what happens in the world. Google calls the speech the real frontier, and it is also where the weak spots show (more on that below). Fourth, movement and physics. DeepMind says people in the simulation move more naturally, and that aerodynamics and collisions look more believable than before. TechCrunch reports Genie 4 holds a consistent scene for up to two minutes, double Genie 3’s limit. DeepMind’s own post is vaguer and gives no number, so treat the two-minute figure as a press-briefing claim. Genie 3 Genie 4 Announced August 2025 27 October 2026 Audio None Generated natively Live text edits Event prompts Weather, camera, people, objects Scene consistency About one minute in Project Genie Up to two minutes (TechCrunch) Public access Project Genie, AI Ultra subscribers Trusted testers only Resolution and frame rate for Genie 4 haven’t been published, so don’t trust any site that quotes a number. How Gemini sits inside Genie 4 This is the part most coverage skips, and it explains why Genie 4 behaves differently. DeepMind says Gemini Omni anchors the simulation in how the real world looks and acts, while a Gemini 3.8 Flash model runs as a guiding agent. It watches incoming frames, splits a goal into steps, checks whether each step worked, and re-plans when you interrupt. So Genie draws the pixels and Gemini keeps the story straight. Google research director Tejas Kulkarni told reporters this kind of pairing, which he called a hybrid world model, will become more common. The clearest demo is robotics. Gemini 4 Argon broke down a task into steps and fed them to Genie 4 as text captions while the scene rendered, simulating Apptronik’s Apollo humanoid playing checkers. The point is to test robot behavior in a simulated room before touching a real one. It’s a research demo, not a product. Who is actually using Genie 4 right now Only trusted testers, and DeepMind has named the groups: filmmakers, photojournalists, educators, game developers and a few enterprise partners. Filmmakers are running “takes” inside a live scene, changing camera angle, lighting and dialogue between runs. Photojournalists are turning a single archival photo into a walkable space. Game developers go from an idea to a prototype they can explore in seconds, before any engine code exists. Diego Rivas of DeepMind put the value in “a state of flow” during ideation and storyboarding. That’s an honest pitch. Nobody at Google is claiming this ships a finished film or a finished game. Can Genie 4 make a complete game?

No, and the official materials never say it can.

What still breaks in Genie 4 DeepMind lists its own limits plainly. Visual details and object permanence can shift over time, especially in longer sessions, after several complex edits, or when you move around a lot. Complex physical interactions, real-time audio accuracy, legible text in the scene, spatial accuracy and long-term consistency are all still open research problems. Read that last sentence slowly if you’re thinking of building anything on it. Signs, labels and subtitles inside the world may come out garbled, and spoken lines may not match what you wanted. The demo reels are also picked by Google, so expect real sessions to look rougher. How to get Genie 4 access and what it will cost Today, Genie 4 itself is not open to the public. The rollout order is trusted testers now, then Google AI Ultra subscribers through Project Genie, then early access partners in the Gemini API and Google AI Studio. Google hasn’t given a date beyond “coming soon.” Here is what you can do: Join the waitlist through the form linked from DeepMind’s announcement. If you want hands-on time right now, Project Genie at labs.google/projectgenie still runs Genie 3, for adult Ultra subscribers, with generations capped at about 60 seconds. Developers should watch the Gemini API and AI Studio announcements. No model ID, pricing or documentation has appeared yet. On cost: Google hasn’t published any Genie 4 price. Trackers report AI Ultra plans between roughly 0 and 0 a month, and sources disagree on which tier includes Project Genie. Check Google’s plan page before paying, and don’t subscribe just to get Genie 4, because subscribing today gets you Genie 3. Google hasn’t published a country list for Genie 4 either, so readers in India should wait for that rather than assume. Why Genie 4 matters beyond the demo Genie 4 landed in a crowded week for world models. TechCrunch notes Runway shipped its second world model a couple of months earlier, and the field is consolidating through acquisitions, including AMD’s reported deal for Fei-Fei Li’s World Labs. Google’s answer is to bolt Gemini’s reasoning onto the simulator, which no rival has matched in public yet. For creators, the practical shift is speed of iteration. You can test a mood, a camera move or a level idea in seconds and throw it away. For the AI field, the bigger bet is robotics and agents, where a simulator that can be steered in language might eventually replace a lot of physical test rigs. Both are still promises, not shipped results. Everything Genie 4 produces, video and audio, carries Google’s SynthID watermark, and DeepMind says it ran internal safety evaluations before opening the preview. Genie 4 questions people keep asking Is Genie 4 available to the public?

No. As of 30 October 2026 it is a research preview for trusted testers, with a waitlist but no public release date.

Google hasn’t announced pricing or a free tier. The consumer route is Project Genie, which currently requires a paid Google AI Ultra plan.

What is a world model in AI?

It’s a model that simulates how a scene changes over time, including physics and how things look and sound, so you can interact with it instead of just watching.

Google Genie 4: How It Works, Features and Who Gets...

बात जो सच तक पहुंचे...