Genie 4 is the new world model from Google DeepMind, announced on 27 October 2026 and opened only to trusted testers. It generates a photorealistic place you can walk through with a keyboard, rewrite mid-session with a typed sentence, and hear as it runs, because the sound is produced by the model along with the pictures. For the first time in the series, the people inside the simulation can talk back.
The demos are striking, and the caveats are unusually plain. By TechCrunch’s account, Genie 4 holds a scene together for up to two minutes. DeepMind says objects and details still drift, audio is not yet reliably accurate and in-world text may not be legible. There is no price, no public model ID and no general release date. This is what the launch shows, what it leaves out, and how it compares with a field that is consolidating fast.
Table of contents
- What Genie 4 Actually Is
- What Changed Between Genie 3 and Genie 4
- Native Audio Is the Headline Genie 4 Feature
- How Gemini Steers Genie 4 From Behind the Scenes
- The Two-Minute Limit Is Murkier Than It Looks
- Where Genie 4 Still Falls Short
- What Trusted Testers Are Doing With Genie 4
- How Genie 4 Compares With Rival World Models
- Who Can Use Genie 4, and When
- Frequently Asked Questions About Genie 4
- References
What Genie 4 Actually Is
A world model, not a video generator
A video generator produces a finished clip and hands it to you. A world model simulates how an environment evolves and how actions change it, then keeps generating the path ahead as you move. DeepMind’s own definition, from its Project Genie launch, is a system that predicts how the dynamics of an environment evolve and how actions affect them.
That difference is why the new model sits alongside Veo in Google’s line-up rather than replacing it. Veo makes high-definition clips with synchronised audio from a prompt. This one is steered while it runs, using standard keyboard controls for navigation and camera, and text prompts for everything else.
Who built it and how it was announced
The launch post is credited to Raia Hadsell, Diego Rivas and Hang Qi, and DeepMind’s thread on X went up at 16:10 UTC on 27 October. The post describes the model as natively multimodal, meaning music, ambient noise and dialogue are generated at the same time as the world rather than added afterwards.
It also frames the work around DeepMind’s stated goal. Understanding and simulating the world, the post says, is a crucial part of building AGI, with safe and infinitely variable simulations useful for training robots and agents as well as for new interactive experiences.
Where it sits in the Genie series
Genie 2 arrived in December 2024. Genie 3 followed on 5 August 2025 as DeepMind’s first real-time interactive world model. In January 2026 that model reached paying users through Project Genie, a research prototype for Google AI Ultra subscribers in the US. The new release comes nearly 15 months after Genie 3, the longest gap in that run.
What Changed Between Genie 3 and Genie 4
Four capabilities, one new foundation
DeepMind groups the upgrade into four areas: real-time shaping of the world, realism in physics and human action, native sound, and complex events reasoned through by Gemini. Underneath them is a dependency that Genie 3 did not advertise. Genie 4 incorporates capabilities from the Gemini family, with Gemini Omni anchoring simulations in real-world appearance and behaviour.
Free-form edits while the world runs
You can now type a request to change the weather, move the camera, or introduce a new person or object. Some coverage treats this as new, but it is more of a refinement. Genie 3 already offered what DeepMind called promptable world events, including altered weather and new objects and characters. The difference is that Gemini now adapts your request to fit the surrounding scene, and between inputs it monitors and steers the world so it keeps progressing.
Physics and people
DeepMind calls its handling of people the most dramatic leap. Characters are meant to move naturally, respond to their surroundings and be steered in real time. On physics, the model is said to handle scenarios such as aerodynamics and collisions, shown in demos of a watermelon on a stone plinth and a falling maple seed.
| Feature | Genie 3 (Aug 2025) | This release (Oct 2026) |
|---|---|---|
| Resolution and frame rate | 720p at 24 fps | Not stated in the launch post |
| Stated duration | “A few minutes” | Up to two minutes, per TechCrunch |
| Native audio | Not a stated feature | Ambient sound, music, dialogue |
| Mid-session editing | Promptable world events | Text prompts adapted by Gemini |
| Gemini’s role | Not part of the model launch | Appearance, planning, frame checks |
| Access at launch | Limited research preview | Research preview for trusted testers |
Native Audio Is the Headline Genie 4 Feature
Sound generated with the scene
DeepMind describes Genie 4 as its first real-time world model to generate audio natively. Every world carries the ambient sound of its own setting, from wind in the trees to waves on a beach, and the sounds in its opening demo reel are, DeepMind says, all produced directly by the model.
That matters because sound is usually a separate step. Keeping it in one generation pass means, in principle, that a wave and the noise of that wave come from the same simulated moment rather than from a track laid over the top.
Dialogue that reacts
DeepMind calls human speech the true frontier. The model introduces interactive dialogue in which characters converse and respond to your inputs or to events in the world, in real time. TechCrunch offers a speculative use from the pitch: stepping into a painting of a historical scene and chatting to its subjects about what happened.
What is still unsolved
The same briefing that sold the feature also limited it. The team told TechCrunch it has not quite worked out how to ensure the audio will be accurate. A dialogue system that sounds natural but says the wrong thing is a bigger problem for a history lesson than for a dragon race, so this caveat decides which uses arrive first.
Runway’s GWM Worlds 2, shared on 3 September 2026, already advertises 48,000 Hz audio, so native sound is not unique to Google. Interactive speech is the sharper claim, and nobody outside the testers has been able to check it yet.
How Gemini Steers Genie 4 From Behind the Scenes
Gemini 3.8 Flash as the guiding agent
This is the most significant architectural change, even though it is the least visible. As Genie 4 generates a world, Gemini 3.8 Flash observes the scene, breaks goals into a sequence of steps, and reviews incoming frames to check that each step succeeded. If the world shifts or you intervene, it changes the plan on the spot.
TechCrunch’s shorthand is that Gemini acts as the director while Genie creates the visuals. DeepMind’s claim is that the arrangement delivers reliable causality, procedural behaviour and multi-step reasoning, which a pixel model alone struggles to sustain.
The robotics experiment with Gemini 4 Argon
DeepMind also tested Gemini 4 Argon with Genie 4 in work tied to Gemini Robotics. In the example it published, Argon breaks a user’s task into steps and feeds them to Genie 4 as text captions while the simulation renders. The scene shows Apptronik’s Apollo humanoid playing a game of checkers.
The appeal is clear enough: test how a robot might handle a scenario before it happens on a real floor. Simulated worlds are also a natural fit for reinforcement learning, where an agent improves through repeated trial and error. Our earlier report on Safeworld and GenAI robot safety covers why that testing matters. DeepMind has published no figures on how well anything learned this way carries over to physical hardware, so treat it as a direction rather than a result.
The case for hybrid world models
DeepMind says future progress lies in hybrid systems that combine the structured reasoning of code and language models with pixel-level simulation. “What we’ll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction,” Tejas Kulkarni, a research director at Google, said in a briefing reported by TechCrunch.
The Two-Minute Limit Is Murkier Than It Looks
One number, one source
The figure most outlets repeated comes from TechCrunch, which reports that Genie 4 can maintain a consistent scene for up to two minutes, double its predecessor’s limit. It does not appear in DeepMind’s launch post. That post says only that the model holds together “over short sessions”, and it gives no duration, resolution or frame rate at all.
Double of what?
Here is the wrinkle. DeepMind’s August 2025 post said Genie 3 retained consistency “for a few minutes” at 720p, and listed its limit as “a few minutes of continuous interaction”. On those words, two minutes is not obviously a doubling. The number it does double is 60 seconds, the session cap that DeepMind placed on Project Genie in January.
The two figures may not measure the same thing. DeepMind’s Shlomi Fruchter told TechCrunch the 60-second cap existed partly because of compute: “when you’re using it, there’s a chip somewhere that’s only yours and it’s being dedicated to your session.” A product cap and a research claim about consistency are different limits, and comparing them flatters or punishes the new model depending on which you pick.
Why the number still matters
Two minutes is enough to block out a shot or test a game mood. It is not enough for a lesson, a training run or a play session. Runway says GWM Worlds 2 sessions “have no preset length”, though that is a statement about the product rather than proof of how long a world stays coherent. Until it reaches independent hands, no one outside these companies can compare them fairly.
Where Genie 4 Still Falls Short
DeepMind’s own list of problems
Credit where it is due: the launch post is frank. DeepMind calls this an early, experimental model, and it names the failures. Visual details and object permanence can shift over time, especially as a simulation runs longer, after several complex prompt edits, or while you navigate around the scene.
| Known gap | What it could mean in use |
|---|---|
| Object permanence drifts | A prop may change or vanish after you turn away |
| Complex physical interactions | Unreliable for testing precise contact or handling |
| Real-time audio accuracy | Speech may sound right and be factually wrong |
| Legible text | Signs and documents in a scene may be garbled |
| Spatial accuracy | Risky for measurement or layout work |
| Long-term consistency | Longer sessions lose earlier details |
The left column is DeepMind’s wording, lightly condensed. The right column is our reading of the practical effect, not a test result, since we have not had access to the model.
No benchmarks to check
DeepMind’s post includes no benchmark table and no head-to-head comparison. DataCamp’s review, checked on 28 October, found no third-party evaluations either. So far, the evidence consists of DeepMind’s own recordings and prompts.
Safety and provenance
DeepMind says the model went through internal safety evaluations before the preview and carries guardrails consistent with its other models. All video and audio generated by Genie 4 will include SynthID, Google’s imperceptible watermark. That is sensible for a tool that can animate a real archival photograph, although a watermark only helps people who actually check for it.
What Trusted Testers Are Doing With Genie 4
Film pre-visualisation
The most concrete early use is not gaming. DeepMind says creatives are using Genie 4 as a rapid “draft mode”: filmmakers can step into a live scene and run several takes, adjusting camera, lighting, dialogue and action as they go. “The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful,” Diego Rivas said at the press briefing, naming ideation and storyboarding.
Archives, games and scenario testing
Photojournalists are exploring how a single 2D archival photograph can become a grounded simulation of that place. Teams are also testing populated historical spaces and short-term events such as a flash flood or wildfire.
Game developers are using it to go from an idea to an explorable prototype in seconds, trying styles and mechanics before writing engine code. That overlaps with what Google is attempting in its Playground game-creation experiment. Enterprise partners, meanwhile, are evaluating rapid “zero-to-one” scenario generation, and DeepMind is candid that they are also identifying gaps.
The honest reading: drafts, not delivery
Every example here is an early-stage use, where a rough result is acceptable and speed is the point. Rivas stressed that it is still early and that the end-to-end product is not yet clear. He also said turning those prototypes into finished assets through Gemini Omni is still some way off.
The sensitive case is photojournalism. Turning a real archive photograph into an inhabited scene means inventing everything the camera did not capture, and DeepMind’s own list of gaps says details will drift. Any newsroom working this way would need to label the result clearly as a simulation.
How Genie 4 Compares With Rival World Models
The closest challenger
Runway’s GWM Worlds 2 is the most direct comparison. It generates interactive simulations in continuous 720p at 24 fps with 48,000 Hz audio, steered with text actions and camera motion, and aimed at entertainment, virtual characters and robotics. Nvidia offers open-weight world models under the Cosmos name, while World Labs sells Marble.
| Model | Owner | Published detail | Status |
|---|---|---|---|
| This release | Google DeepMind | Native audio, live edits, Gemini planning | Trusted testers |
| GWM Worlds 2 | Runway | 720p, 24 fps, 48,000 Hz audio | Shared Sept 2026 |
| Marble | World Labs | Entertainment and robot training worlds | Commercial; AMD deal pending |
| Oasis | Decart | Simulates real-world 3D environments | Amazon talks reported |
| Cosmos | Nvidia | Open-weight world models | Available |
A market consolidating around simulation
Genie 4 lands as independents get bought. AMD agreed to buy World Labs for $8.2 billion in September. The Wall Street Journal reported that Amazon is in talks to buy Decart for around $7 billion, against a valuation of nearly $4 billion in May, and TechCrunch still described the two as in talks on 27 October. Rivian bought Third Dimension, a 15-person UK startup, on undisclosed terms.
Why buyers want this
The buyers’ motives are practical rather than playful. Rivian wants to turn real driving footage into editable simulations for validating self-driving software, a job where trust in the gap between simulation and reality decides whether a system ships. TechCrunch notes that scarce real-world data makes synthetic data important for training general-purpose robots.
Startups are also taking the idea into specific industries, as Dream’s construction-focused world models show. Against that backdrop, Google has an advantage few rivals can copy: it can wire its world model straight into Gemini. Our AI models and tools hub tracks how these releases keep shifting.
Who Can Use Genie 4, and When
The rollout order
Access comes in stages, and only the first is open. DeepMind says the model is coming “soon” to the next groups, and TechCrunch reports “in the coming weeks”. Neither gives a date, so treat any specific day you see elsewhere with suspicion.
| Group | Route | Status at 31 Oct 2026 |
|---|---|---|
| Trusted testers | DeepMind cohorts | Live now |
| Google AI Ultra subscribers | Project Genie | Coming soon, waitlist open |
| Early access developers | Gemini API and Google AI Studio | Coming soon, partners only |
| Everyone else | None announced | No date given |
What has not been published
There is no pricing, no per-session cost and, per DataCamp’s 28 October check, no public model ID. That is a real gap, because a chip dedicated to each session was the stated reason for Project Genie’s 60-second limit in January. Nothing yet tells us what a Genie 4 session costs to run, or how long paying users will get.
What to do now
If you make films, games or training content, join the Project Genie waitlist and plan a small test on rough-draft work. Project Genie launched for US subscribers aged 18 and over, and DeepMind’s launch post does not say which countries will get this model. If you run a robotics or simulation team, wait for the API terms, because a demo of a humanoid playing checkers tells you little about how far results transfer to hardware.
Frequently Asked Questions About Genie 4
What is Genie 4?
It is Google DeepMind’s latest world model, announced on 27 October 2026. It generates interactive, photorealistic environments in real time with native audio, which you steer with keyboard controls and text prompts.
Can I use Genie 4 today?
Only if you are one of DeepMind’s trusted testers. Google AI Ultra subscribers using Project Genie and early access Gemini API and Google AI Studio partners are next, and there is a public waitlist. No date has been given.
How much does Genie 4 cost?
No price has been published. Access for consumers is expected to arrive through a Google AI Ultra subscription, but DeepMind has not announced session limits or API rates.
How long does a Genie 4 world last?
TechCrunch reports up to two minutes of consistent scene. DeepMind’s own post gives no number and refers only to short sessions, warning that details drift over longer runs.
Is Genie 4 better than Runway’s GWM Worlds 2?
Nobody independent can say yet. DeepMind published no benchmarks and the model is not publicly available. Runway has published resolution, frame rate and audio figures, while DeepMind’s launch post has not.
Is generated Genie 4 content watermarked?
Yes, according to DeepMind. All video and audio generated by Genie 4 will include SynthID, an imperceptible watermark used to identify AI-generated content.
References
Genie 4: A new generation of world simulation
Google’s latest world model adds audio, more realistic worlds, and real-time steering
Google DeepMind’s launch thread on X
Genie 3: A new frontier for world models
Project Genie: Experimenting with infinite, interactive worlds
I built marshmallow castles in Google’s new AI world generator
AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion
Rivian buys world model startup to accelerate self-driving ambitions
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
