Genie 4 vs Genie 3, side by side
The same dressing-room scene generated by both models. The comparison is the single clearest measure of what one year changed.
Source: Google DeepMind, Genie 4 announcement
Genie 4 turns a sentence into a world you can walk through
Genie 4 is Google DeepMind's real-time, natively multimodal world model, announced on October 27, 2026.
This is an independent guide to what it does, how it differs from Genie 3, and how to actually use it.
Independent guide · not affiliated with Google. Every factual claim is linked to the official DeepMind announcement.
The short answer
Genie 4 is a world model, not a video generator. Instead of rendering a fixed clip, it builds a playable environment you can steer in real time with plain language — complete with physics, ambient audio, and characters that answer you.
- Released — October 27, 2026, in a DeepMind research preview for trusted testers.
- Architecture — Built on Gemini: Gemini Omni grounds the visuals, Gemini 3.8 Flash runs a real-time guiding loop.
- Input and output — Natively multimodal — image, audio, and dialogue are generated together, not stitched afterwards.
- Steering — Change weather, move the camera, or add a character mid-session without restarting the scene.
- Access — Trusted testers first, then AI Ultra subscribers via Project Genie, plus Gemini API and AI Studio partners.
- Watermark — All generated video and audio carry SynthID, an imperceptible watermark.
Google has not published frame rate, output resolution, model size, or pricing for Genie 4. Any exact number you see elsewhere is unverified.
Watch Genie 4 actually run
These are real demo clips published by Google DeepMind alongside the Genie 4 announcement — not illustrations. Every card plays the original footage.
Silk scarf in wind
A typed prompt — "wind blows on silk" — drives a pink ribbon across a meadow. The reasoning panel on the right shows the agent resolving the rule.
Source: Google DeepMind, Genie 4 announcement
Marble bust on a watermelon
A heavy object is dropped onto a soft one. A direct look at contact physics and how far the simulation will follow the consequence.
Source: Google DeepMind, Genie 4 announcement
Maple seed in grass
Macro-scale detail: a seed among blades of grass, with light and occlusion at a scale that usually breaks generated worlds.
Source: Google DeepMind, Genie 4 announcement
First-person magician
A wand, a cat and a wooden chest in an interior. Interesting for what the model keeps consistent as the viewpoint moves.
Source: Google DeepMind, Genie 4 announcement
Island piano duet
A character performs on a shoreline. Genie 4 generates audio with the scene, so this is one of the clips where sound matters as much as picture.
Source: Google DeepMind, Genie 4 announcement
Crowned crane on snow
From the opening film. The official caption notes results are illustrative and sequences shortened — worth remembering when judging fidelity.
Source: Google DeepMind, Genie 4 announcement
Laboratory robotics task
A robotic arm moves a pipette over a well plate while the reasoning panel plans the next step. This is the agent-training use case made concrete.
Source: Google DeepMind, Genie 4 announcement
Header motion study
The short loop DeepMind used at the top of the announcement post — a compact look at how the world holds together frame to frame.
Source: Google DeepMind, Genie 4 announcement
Clips are the original files published by Google DeepMind, re-hosted for reliability. genie4.art is not affiliated with Google; all rights remain with the original publisher. Some clips carry DeepMind's own caption that sequences were shortened for illustration.
Genie 4 at a glance, against Genie 3
Genie 3 introduced real-time interaction in 2025. Genie 4 extends it from a single generated scene into a multimodal, agent-driven simulation.
| Dimension | Genie 4 | Genie 3 |
|---|---|---|
| Release | October 27, 2026 | August 5, 2025 |
| Modality | Natively multimodal: visuals, ambient audio, music, live dialogue | Visual frames with promptable world events |
| Reasoning layer | Gemini Omni grounding plus a Gemini 3.8 Flash agent loop that observes, plans, and replans | Not described as agent-driven |
| Interaction | Real-time natural-language steering of weather, camera, objects and people | Real-time navigation plus text-triggered world events |
| Audio | Native ambience plus interactive character dialogue | No native audio generation described |
| Consistency | Longer coherent sessions, though Google still reports drift over long runs | Largely consistent for several minutes; visual memory around one minute |
| Access | Trusted testers, AI Ultra via Project Genie, Gemini API and AI Studio partners | Limited research preview for a small cohort of academics and creators |
Model specs Google has not published — frame rate, resolution, session length — are deliberately left blank rather than estimated.
Scroll the table sideways to see all columns.
How it works
Three layers do the work
DeepMind describes Genie 4 as a Gemini-based system rather than a single monolithic model. Understanding the split explains both its strengths and its failure modes.
1
Gemini grounds reality
Gemini Omni anchors the generated world in real-world appearance and behaviour, so scenes read as places rather than texture.
2
An agent keeps it coherent
Gemini 3.8 Flash runs a real-time loop: observe the frame, split the goal into steps, check whether it worked, replan when you intervene.
3
Rendering makes it playable
The world is recomputed continuously, so your input changes the scene instead of sitting in a queue behind a fixed render.
Four capabilities that define Genie 4
These are the areas DeepMind highlights in the announcement — and the ones that separate a world model from a text-to-video tool.
Real-time steering
Change weather, reposition the camera, drop in a new object or person. Gemini rewrites the prompt to fit context and keeps the scene running between edits.
Physics and human realism
Aerodynamics, collisions, and natural character movement with reactions — the parts that make an environment feel solid.
Native audio
Setting-appropriate ambience plus interactive character dialogue that adapts live, generated with the scene rather than dubbed over it.
Complex event simulation
Multi-step reasoning through Gemini to model causality and procedural behaviour — the basis for robotics and agent training.
SynthID watermarking
Every generated frame and audio track carries an imperceptible watermark, so provenance survives re-encoding.
Promptable world events
Inject scripted moments into a live session — a storm rolls in, a door opens, a crowd arrives — without cutting away.
Where Genie 4 is already being tested
DeepMind names the following workflows, drawing on what trusted testers reported during the research preview.
Film and TV pre-visualisation
Testers describe a "draft mode" for scenes: block a shot in a live generated set, try a take, and iterate before committing budget to a physical build.
Game prototyping before engine work
Explore whether a level layout, camera behaviour, or mechanic feels right while it is still an idea — before writing engine code or building assets.
Robotics and agent training
Generate environments with causal, multi-step events so embodied agents and robots can practise rare scenarios without hardware or a physical set.
Bringing archival images to life
Testers turned 2D historical photographs into immersive simulations — a way to walk through a documented moment instead of looking at a still.
Frames from the official demos
Stills lifted directly from the clips above — genuine Genie 4 output, not illustration. Open any frame full size, or watch the source clip in the gallery above.
Crowned crane, snowfield
Silk scarf in a meadow
Marble bust over a watermelon
Maple seed among grass
Interior with wand and object
Laboratory manipulation task
Shoreline performance
Header motion study
Can you use Genie 4 today?
Access is staged. Here is what DeepMind has confirmed, and what is still unknown.
The first wave. Filmmakers, studios, researchers and robotics teams were given access before any public release to report on what worked and what broke.
Genie 4 questions, answered directly
The questions people actually search for, with answers drawn from the official announcement.
More questions answered on the full FAQ page.
Start with a world prompt
Draft the scene here, then take it to a generator that can actually run it.
No account needed on this site — it is a guide, not a service.