What did Google DeepMind announce on October 27, 2026?
Google DeepMind announced Genie 4 on October 27, 2026. It is the successor to Genie 3, which the lab released as a research preview in August 2025. The new model generates photorealistic, interactive scenes that respond to keyboard controls and typed instructions.
A world model is not a chatbot that predicts the next word. Google defines one as a system that "simulates the dynamics of an environment, predicting how they evolve and how actions affect them." Genie 4 renders a place and keeps it running. It reacts when a user moves the camera, types a change or speaks to a character.
World models are often confused with video generators. Google's Veo makes high-definition clips, with synchronized audio, from a text or image prompt. Genie is different because its output is designed to be steered while it runs. The user's next input changes what happens instead of starting a new clip.
DeepMind lists four changes. Worlds look more realistic, with better physics and more natural human movement. Users can edit a scene in real time using plain language. The model generates sound natively. And Gemini models now help plan and check multi-step events inside a simulation.
The Gemini family sits underneath the model. Google says Gemini Omni anchors the simulations "in real-world appearance and behavior." Other Gemini models supply the reasoning that keeps a scene coherent. DeepMind argues that understanding and simulating the world is "a crucial aspect of building AGI," which is why the work sits at the center of its research agenda.
Availability is narrow. Genie 4 is a research preview for trusted testers, with a public waitlist. Google says it is coming soon to Google AI Ultra subscribers in Project Genie. Early-access partners on the Gemini API and Google AI Studio are next in line. CNET reported that Genie 4 is not available to the general public.
How is Genie 4 different from Genie 3?
Genie 3 set the baseline in August 2025. It generated navigable worlds at 720p and 24 frames per second. DeepMind said it retained consistency "for a few minutes," with visual memory reaching back about one minute. Its output was silent: DeepMind calls Genie 4 its first real-time world model to generate audio natively.
Sound is the most visible addition. Every Genie 4 world carries ambient noise, from wind in trees to waves on a beach. Characters can hold interactive dialogue that adapts to events. Google says the demo audio was produced by the model itself, not added in post-production.
Control is the second change. Users move through a scene with standard keyboard controls and steer it with text. They can change the weather, move the camera, or introduce a new person or object. Between prompts, Gemini watches the scene and keeps it progressing, according to DeepMind.
The third change is realism. DeepMind says the largest step is in how Genie 4 simulates people. Characters move naturally, respond to their surroundings and can be steered in real time. The model also simulates aerodynamics, collisions and other physical behavior.
Duration is the fourth. TechCrunch and The Verge both report that Genie 4 holds a scene together for about two minutes. That is roughly double Genie 3's limit. DeepMind's own Genie 3 post had described environments that stayed largely consistent "for several minutes."
That leaves an inconsistency worth flagging. The two-minute figure comes from reporters briefed by Google, not from a published specification. DeepMind has not disclosed Genie 4's resolution, frame rate or session limits. Treat the duration claim as a working estimate until independent tests appear.
There is a technical reason the limit matters. Genie 3 built worlds frame by frame from the prompt and the user's actions, rather than storing an explicit 3D scene. DeepMind said that approach produced richer, more dynamic worlds. It also meant small errors accumulated over time. Genie 4 inherits the same trade-off.
What can you actually do inside a Genie 4 world?
Google's demos show the range it is aiming for. A saxophonist plays on a floating concrete platform above a lake. A hot-air balloon drifts over Antarctic glaciers. A dragon race unfolds in first person. A crowded 1980s Japanese train station stays busy as the camera moves through it.
The prompts behind those clips are long and specific. Users describe lighting, film grain, character behavior and camera movement in detail, then adjust the scene live. The model fills in how objects collide, how fabric moves and how people react to what happens around them.
DeepMind says its testers are using the model in four broad ways. Filmmakers are previsualizing scenes and running quick takes. Photojournalists are turning archival photographs into explorable places. Game developers are prototyping ideas before writing engine code. Enterprise partners are generating scenarios for simulation work.
The audio is not a soundtrack added afterwards. DeepMind says ambience, music-like sound and dialogue are generated alongside the world itself. That matters for immersion, and potentially for robot training, where sound can carry information about events in a scene.
Diego Rivas, a Google DeepMind engineer, said real-time control matters most when it helps users keep "a state of flow." For him, that points to storyboarding and early ideation rather than final production. The value is speed of iteration, not polished output.
These are Google-selected examples from a press briefing, not independent trials. The Verge's reporter said he had not been able to try Genie 4. He could not compare it with the Genie 3-powered Project Genie. The coverage published so far relies on Google's demos and a press briefing.
Why is Gemini part of the Genie 4 loop?
Genie 4 does not work alone. Gemini 3.8 Flash acts as a guiding agent inside the loop. It observes the generated frames, breaks a goal into steps and reviews progress. When the user intervenes or the world shifts, it revises the plan.
That division of labor is what makes longer sequences possible. A single prompt can set an event in motion. The Gemini layer then keeps the world on track while Genie renders pixels, sound and motion.
DeepMind also ran an experiment that combined Gemini 4 Argon, Genie 4 and Gemini Robotics. Argon split a task into steps and passed them to Genie as captions. The simulation showed an Apptronik Apollo humanoid playing checkers. The goal was to test whether a robot task can be rehearsed in simulation first.
Tejas Kulkarni, a research director at Google, described the direction as hybrid world models: "a combination of Gemini writing code and text, and Genie ingesting that in real time." The idea is to pair language-level planning with pixel-level physics.
Language models are good at structure and sequence. World models are good at appearance and motion. Google argues that each covers the other's weak spot, though the checkers demo is a research illustration rather than a shipping feature. DeepMind calls hybrid world models future work, not a finished product.
The framing is broader than one model. Google says the combination lets teams test how AI systems navigate different environments before bringing them to the real world. A model that can check whether a step succeeded is also a model that can score an agent's behavior without a human watching every frame.
Rivas also said users could eventually use Gemini Omni, Google's multimodal model, to turn those prototypes into finished assets. He added that the capability is still some way off. It is also not yet clear, he said, what the end-to-end product experience will be.
Why do world models matter outside games?
The commercial pitches overlap: games, film, robotics and training. The robotics case is the most consequential. Robots need experience, and real-world trials are slow, costly and hard to repeat.
A world model offers a cheaper place to practice. DeepMind says Genie 4 can create "safe, infinitely variable simulations" for training robots and agents. Testers are also using it to visualize short-lived events such as flash floods and wildfires. Those are hard to stage in reality.
Money is moving in the same direction. AMD announced on September 28, 2026, that it will buy World Labs, the spatial-intelligence lab led by Fei-Fei Li. The all-stock deal is valued at about $8.2 billion and is expected to close by the end of 2026. Li will become AMD's executive vice president and chief scientist.
"Joining AMD will give our team the resources and engineering depth to accelerate our research," Li said. AMD says the purchase will help it understand how AI workloads are changing as the field moves into reasoning, robotics and physical AI. World Labs builds models that generate and simulate interactive 3D environments, plus technology for robot learning.
Google is not alone on the model side. Runway released GWM Worlds 2 on September 3, 2026. It is a playable world model that runs at 720p and 24 frames per second with 48,000 Hz audio. Runway says its sessions have no preset length, which sets a higher bar than Google's two minutes on paper.
Runway describes GWM Worlds 2 as an extension of its GWM Worlds research from December 2025. Users can address text actions to any subject or to the scene itself, alongside continuous camera movement. The company is selling control and persistence, not photorealism alone.
Each company is betting on a different strength. Google has the Gemini stack and a consumer distribution channel through Project Genie and AI Ultra. Runway is pitching continuous sessions and scene control. AMD is buying model expertise to shape future chips. The competition is now about endurance and control, not just image quality.
What are Genie 4's biggest limitations?
Duration is the obvious one. Two minutes is enough for a demo and for some pre-visualisation work. It is not enough for a film scene, a long game level or an extended robot training run.
Consistency is the deeper problem. DeepMind says visual details and object permanence can shift as a simulation runs longer. Repeated prompt edits and camera movement make it likelier. The world can look convincing while quietly changing behind the user.
Audio and text are unfinished too. The company lists real-time audio accuracy and legible in-world text as open research problems. It puts complex physical interactions and spatial accuracy in the same category. In practice, the model can produce convincing sound and signage without guaranteeing either is correct.
Every output carries SynthID, Google's imperceptible watermark, across both video and audio. DeepMind says it ran internal safety evaluations before the preview and limited access to trusted testers first. Those steps reduce some misuse risk, but they do not make the output a verified record of anything.
The risk is not only technical. CNET noted that public anxiety about AI has shifted toward job losses, misinformation and deepfakes. A model that can convincingly restage a street or a person makes provenance questions more pressing. That is why Google is pushing watermarking alongside the model.
The biggest gap for buyers is evidence. Google has published curated clips, not benchmarks. There is no public measure of physics accuracy or audio fidelity. Nor is there evidence on how well robot policies trained inside Genie 4 transfer to real hardware.
DeepMind's own framing is cautious. The blog calls Genie 4 "an early, evolving area of research" and says the preview exists to learn from users. Google has not said how many testers it has, which partners are involved, or what feedback would trigger a wider release.
When can you try Genie 4, and what should you watch next?
The waitlist is open now. Google says Genie 4 is coming soon to AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and AI Studio. TechCrunch reported the Ultra rollout would happen in the coming weeks. Google has not given a firm date.
A research preview also means no service guarantees. Google can pause access, change the model or restrict outputs while it gathers feedback. The waitlist gives it a way to control who tests what, and at what scale.
Project Genie opened on January 29, 2026, to AI Ultra subscribers in the United States aged 18 and over. It runs on Genie 3. The prototype creates worlds from text and images, and lets users explore and remix them. Google has not said whether Genie 4 will replace Genie 3 inside Project Genie or sit alongside it.
Pricing remains unclear. Google has not announced a standalone price for Genie 4, and Project Genie requires an AI Ultra subscription. The company has also not published rates for API access, which will matter more to developers than the consumer waitlist.
Three things are worth watching. First, whether independent testers confirm the two-minute limit and the audio quality. Second, whether Google discloses resolution and frame rate, which it has not done for Genie 4. Third, whether simulated robot training produces measurable gains on real machines rather than just impressive clips.
For developers, the API route will matter most. Session length, cost per simulated minute and output resolution will decide whether Genie 4 becomes a production tool or stays a demo.
For now, Genie 4 is a research preview with strong demos and no public benchmarks. That is a normal stage for world models. It is also a reason to treat the videos as evidence of direction, not proof of capability.