Google DeepMind has released Genie 4, a real-time world model capable of generating, interacting, and accepting new commands simultaneously. Users can not only move within the generated environment but also instantly alter weather, camera angles, characters, and events using natural language; the model synchronously generates ambient sounds, music, and interactive character dialogue. This is not achieved by a single standalone model: Gemini Omni helps create more realistic appearances and behaviors, Gemini 3.8 Flash continuously observes the generated visuals, breaks down goals, and verifies each step, while Genie 4 renders the world frame by frame. Google positions it as infrastructure for film pre-visualization, game prototyping, robotics, and AI agent training environments. However, Genie 4 is currently only available to invited testers, continuous interactions last only a few minutes, objects may change once they leave the frame, and complex physics, spatial relationships, text, and audio are not yet fully reliable; the company has not disclosed parameters, training data, computational costs, standardized evaluations, or API pricing.Article author, source: Google DeepMind
It generates not a fixed video, but a scene that continues to evolve.
Traditional text-to-video models, after receiving a prompt, typically output a pre-determined video. Once generation is complete, if users want to change the weather, add characters, or modify the plot, they often need to regenerate the entire video.
Genie 4 uses a different interaction method. Users can control movement and camera angles with a standard keyboard while continuing to type text during world generation—for example, requesting that it start raining, moving the camera, adding a new character, or prompting characters in the scene to react differently. Gemini interprets the new instructions in the context of the current scene, and Genie 4 continues generating subsequent visuals and audio.
This means the model must remember what it has generated previously and predict what should happen after each action. It no longer just deals with “what the next frame looks like,” but also how characters move, how objects collide, how the environment responds to the user, and how sounds change with events.
First time adding native audio and live person-to-person dialogue
Genie 4 is Google DeepMind's first native generative audio real-world model.
The model generates sounds such as wind, ocean waves, footsteps, vehicles, musical instruments, and indoor echoes based on the current environment, rather than adding a separate audio track after the visuals are completed. Characters can also engage in real-time dialogue based on user input and events within the scene.
For example, in the dragon-riding demo, the system synchronously generates the sound of flapping wings, wind at high altitude, flames, and distant volcanic activity; in a motorcycle scene modeled after 1990s Hong Kong films, the model simultaneously processes character movements, vehicle motion, road environment, and dialogue.
This differs from Google’s Veo video model. Veo primarily outputs highly polished film clips, while Genie emphasizes user interaction within the scene and the ability to manipulate and change the world. The former is more like a finished shot, while the latter is closer to a real-time simulation.
However, "native audio" does not guarantee consistent accuracy. Google explicitly lists real-time audio accuracy as an unresolved issue; community testers have also noted that in some demonstrations, prompts requesting a saxophone sound produced results closer to a clarinet.
A truly new structure, where Gemini acts as the "director" and Genie as the "cinematographer".
Genie 4 does not independently maintain the entire simulation.
Gemini Omni helps the model develop more realistic appearance and behavior. When faced with multi-step tasks, Gemini 3.8 Flash acts as a guiding agent within the loop: continuously observing the visuals generated by Genie, breaking down the user’s goal into steps, verifying whether the current step is complete, and readjusting the plan based on user input.
Genie 4 translates these plans into real-time visual, audio, and physical feedback. Think of it as a simple division of labor: Gemini handles structure, causality, and what should happen next; Genie renders the immediate world pixel by pixel.
Google also demonstrated a robotics experiment combining Gemini 4 Argon with Genie 4. Argon first breaks down the task into action steps and continuously sends textual instructions to Genie, which simulates the Apptronik Apollo humanoid robot to complete the checkers task.
This is not a robot that has already learned to play chess in the real world. The demonstration took place in a generative simulation environment, whose value lies in allowing the agent to attempt a wide range of tasks and edge cases without damaging real equipment or requiring manual setup of simulation scenarios for each situation.
Google wants to turn world models into training grounds for robots and agents.
Movies and games are the most intuitive uses for Genie 4.
Film production teams can use it as a real-time previsualization tool: enter a scene and continuously adjust lighting, character positions, camera angles, and dialogue without first completing formal modeling, filming, or post-production. Game developers can directly obtain an exploratory concept prototype from a single description, testing visual style, materials, and interaction mechanics before writing any code.
Google also mentioned two less prominent uses: news and education teams can expand a historical photograph into an exploratory environment; businesses can generate short-duration events such as fires, floods, or mechanical accidents to discuss emergency procedures.
The long-term goal is to train robots and general-purpose agents. Real-world robotic data is expensive, slow, and may pose safety risks; traditional simulators require humans to manually build assets, physical rules, and tasks. If sufficiently stable, generative world models can automatically create vast numbers of diverse rooms, roads, factories, and unexpected scenarios in which agents can learn.
Therefore, the significance of Genie 4 lies not just in “more realistic visuals,” but in Google beginning to combine the planning capabilities of language models with pixel-level world generation into a hybrid system: logic is maintained by Gemini, while perception of the world is generated by Genie.
The biggest issue remains object permanence: when out of frame, the world may change silently.
Genie 4 is still far from being a true game engine or a reliable robot simulator.
The official acknowledgment states that visual details and object permanence may drift over time, especially after multiple prompt modifications, long-distance movements, or when objects leave the frame. Hacker News users pointed out specific changes in the official demo: passengers walking away in a train station changed their appearance; a castle appeared in the distance in a scene featuring a rider on a dragon; and the motorcycle rider’s helmet state differed before and after.
These phenomena indicate that while the model can generate visually coherent local scenes, it may not maintain a consistent underlying 3D world state. Traditional game engines track the position, properties, and collision relationships of every object; generative models, however, may reinterpret what an object should look like each time they encounter it, based solely on the current context.
This discrepancy may be acceptable for film pre-visualization, but it is critical for robot training. If the cup in the simulated environment changes position after being out of view, or if collision rules vary each time they are run, the strategies the robot learns may not transfer to the real world.
Other limitations listed by Google include: complex interactions among multiple users and independent agents are still inaccurate; real-world locations cannot be fully reproduced; text within scenes is difficult to generate stably; spatial relationships and complex physics may be incorrect; continuous interactions can only be maintained for a few minutes; and audio and long-term consistency still require improvement.
It's not yet an "infinite game" open to everyone.
Genie 4 is initially available to invited testers in creative, news, education, gaming, and engineering fields. Google stated that it will later roll out to Project Genie users on Google AI Ultra, as well as early partners of the Gemini API and Google AI Studio, but did not disclose specific dates, usage limits, or pricing.
It is also not an open-weight model. The official team has not disclosed parameters, training data, training compute, resolution, frame rate, inference hardware, or cost per minute of generation. Currently, external parties cannot independently deploy it or assess the computational resources required to achieve the demonstrated performance.
Google states that all videos and audio generated by Genie 4 will include an invisible SynthID watermark and undergo internal security evaluations before broader deployment. However, the company has not publicly released the full security assessment, false positive rates, or results of bypass tests.
The most significant change in Genie 4 is that generative AI is shifting from “producing a piece of content” to “maintaining an environment capable of ongoing action.” While it has not yet established a stable, persistent, and verifiable physical world, it has already integrated video generation, audio generation, language planning, and real-time interaction into a single loop.