Technology & Research · 2026-10-28 · Team LATENT

Can Genie 4 replace a game engine? How Gemini watches over a generated world, and where it falls short

Google DeepMind's Genie 4 is a world model that generates video and sound together in response to key presses and text instructions. We read the official description of how Gemini 3.8 Flash watches the scene and steers it, then compare Genie 4 with game engines, which keep world state as numbers, and sort out what game developers can try with it.

Apptronik's Apollo humanoid robot moves pieces in front of a checkers board. In this video published by Google DeepMind, the robot, the board, and the pieces are all generated pixel by pixel by Genie 4. Gemini 4 Argon breaks the task the user asked for into steps. Argon passes each step to Genie 4 as text, and Genie 4 keeps rendering the next scene according to that text.

On October 27, 2026, Google DeepMind announced Genie 4, a world model. When a user moves and controls the camera with a keyboard or similar input and adds instructions in text, Genie 4 generates the next frames and sound in response. It is currently a research preview for trusted testers, and the general public and developers cannot use it yet.

Game worlds are usually built in game engines such as Unity, Unreal Engine, and Godot. If you can generate a world from text and control it, do you still need a game engine? This article reads the official description of how Gemini steers Genie 4, then compares how Genie 4 and game engines keep the state of a world. Specifications were checked on October 28. Statements about Genie 4's performance come from DeepMind's descriptions and published videos, and LATENT has not used the model. The comparison with game engines is LATENT's analysis.

Google DeepMind announcement / Genie 4 model page

Key takeaways

Genie 4 generates video together with ambient sound, music, and dialogue in response to controls and text instructions. For complex events, Gemini 3.8 Flash watches the scene, breaks a goal into steps, and checks in the generated frames whether each step succeeded.
Game engines keep object positions and collision shapes as numbers and run rules as code. Genie 4 generates the look of the world as pixels, while Gemini handles structure and planning. DeepMind lists object permanence, text rendering, and interaction limited to a few minutes among its limitations.
Genie 4 is set to expand to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio. Resolution, frame rate, pricing, and a launch date have not been published.

Genie 4 generates video and sound together in response to controls

According to DeepMind, Genie 4 incorporates and builds on capabilities from the Gemini family. Gemini Omni keeps scenes consistent with real-world appearance and behavior, and other Gemini models provide the reasoning needed to keep a scene coherent and continuous. Gemini Omni is a model that generates video from image, text, video, and audio references and lets you edit it through conversation.

DeepMind describes Genie 4's capabilities in four parts. The first is changing a scene with natural language. Users can change the weather, move the camera, and add new people or objects. Gemini adjusts the user's instruction to fit the surrounding scene. Between instructions, Gemini also monitors the scene and steers the world so that it keeps progressing coherently.

The second is realistic physics and human motion. DeepMind says Genie 4 simulates physical phenomena such as aerodynamics and collisions, and names its simulation of people as the biggest leap. Generated characters move naturally, respond to their surroundings, and can be steered in real time, DeepMind says.

The third is sound. DeepMind describes Genie 4 as its first real-time world model to generate audio natively. Each world carries ambient sound for its setting, such as wind in the trees or waves on a beach. Characters speak in response to the user's input and to events in the world. The fourth is simulating complex events in combination with Gemini, covered in the third section.

DeepMind says the published videos are recordings of real-time interaction. Users move and control the camera with a keyboard and mouse and steer the scene with text. The model page's disclaimer, however, says the results are for illustrative purposes and the sequences have been shortened. No quantitative evaluation of the physics or human motion has been published.

Google DeepMind announcement / Genie 4 model page / Gemini Omni model page

What changed from Genie 3, and what has not been published

The previous model, Genie 3, was announced in August 2025. The table below lines up what DeepMind's published materials say about the two models.

Item	Genie 3 (August 2025)	Genie 4 (October 2026)
Resolution and frame rate	720p, 24 frames per second	Not published
Consistency	Largely consistent for several minutes; visual memory reaches back one minute	Coherent over short sessions; details and objects drift over time
Continuous interaction	A few minutes	A few minutes
Sound	None	Ambient sound, music, and dialogue generated together
Changes by text	Text can trigger events in the world	Change weather, camera, people, and objects; Gemini adjusts instructions to the scene
Handling of goals	The model is not aware of the agent's goal	Gemini 3.8 Flash breaks goals into steps and checks progress

Genie 3 was announced as generating worlds from text prompts at 720p and 24 frames per second. Neither the Genie 4 announcement nor its model page gives a resolution or frame rate. On the length of continuous interaction, the Genie 4 model page says a few minutes rather than extended sessions. That is nearly the same wording as the Genie 3 announcement, so the official materials do not show whether the interaction time has grown.

TechCrunch reported that Genie 4 can maintain a consistent scene for up to two minutes, double its predecessor's limit. Project Genie, the prototype app Google launched on Genie 3 in January, limited generations to 60 seconds. The TechCrunch article does not say what the doubling was measured against. We could not find the two-minute figure in DeepMind's published materials.

Genie 3 announcement / Genie 4 model page / TechCrunch article

Gemini 3.8 Flash watches the scene and checks whether each step succeeded

To handle complex events in real time, Genie 4 brings in Gemini 3.8 Flash as a guiding agent in the loop. In DeepMind's description, Gemini keeps observing the scene while Genie 4 generates the world. It breaks a goal into a sequence of steps, reviews incoming frames to track progress, and verifies whether each step succeeded. When the world shifts or the user intervenes, Gemini revises its plan on the spot. DeepMind says this combination enables worlds with reliable causality, procedural behaviors, and multi-step reasoning.

With Genie 3, the relationship between the model and the agent was different. DeepMind ran its SIMA agent for 3D environments in worlds generated by Genie 3, giving it goals to pursue. DeepMind explained that Genie 3 was not aware of the agent's goal and only simulated the future based on the agent's actions. In Genie 4, Gemini knows the goal and steers progress from outside the world.

In the robot example at the top, Gemini 4 Argon plays the same role. Argon breaks the user's task into steps and interleaves them as text captions while the simulation renders. Because Genie 4 can be steered with text, the simulation can be adjusted along the way toward completing the task. LATENT covered Argon on October 2 as a model with an output limit of 1 million tokens. At that time, early access to Argon was limited to a group of defenders.

In this setup, whether a step succeeded is also judged from video. In a game engine, a program can check from coordinates whether a piece landed on the correct square. With Genie 4 and Gemini, Gemini looks at the generated frames and decides. How accurate that judgment is depends both on how accurately Gemini reads the frames and on how faithfully Genie 4 renders the instruction. DeepMind has not published figures for how accurately it verifies steps. For the checkers example, it also has not shown any check of whether the moves followed the rules.

Google DeepMind announcement / Genie 3 announcement / Related: Gemini 4 Argon takes on long-running work with up to 1 million output tokens

Game engines keep state as numbers; Genie 4 generates it as video

A game engine keeps the objects in a scene as numerical data. According to Godot's introduction to physics, to detect contact between objects you assign at least one collision shape to each object. You apply forces such as gravity or impulses to a rigid body, and the physics engine calculates the resulting movement. Game rules are written as code, and state such as score or inventory is stored in variables. If you build a save system, you can resume from the same scene, and you can check values in tests.

Physics in a game engine does not always return the same result either. The same Godot page warns that Godot's physics is not deterministic and is not guaranteed to behave the same way in situations that look identical. Even so, an object's position and whether a collision occurred can be read out as numbers.

DeepMind described a different approach for Genie 3. According to the Genie 3 announcement, NeRFs and Gaussian Splatting depend on an explicit 3D representation. Genie 3, by contrast, creates the world frame by frame from the world description and the user's actions. When generating each frame, the model refers to the trajectory it has generated so far. If the user returns to a location after a minute, the model has to refer back to the information from a minute earlier. DeepMind explains that generating an environment auto-regressively like this is harder than generating a whole video, because inaccuracies tend to accumulate over time.

Genie 4's architecture has not been published. As an early step toward future work, DeepMind describes how Gemini and Genie divide the work. Gemini works behind the scenes to map the structure of the world and plan the narrative timeline. Genie generates how the world looks, sounds, and responds, pixel by pixel.

A world that looks natural is not necessarily accurate as a game. A game has to decide by its rules whether a bullet hit an enemy or whether the player holds a door key, and it has to remember the result. In a game engine, collision checks and variable values decide these things. In Genie 4, whether something looks like a hit is decided by the generated frames. Whether any state can be read out as numbers has not been published.

The table below lines up the limitations DeepMind lists on its model page with the game features they correspond to. The right-hand column is LATENT's mapping.

Limitation listed by DeepMind	Corresponding game feature
Details and object permanence drift over time	Items you placed and enemies you defeated stay where they are
Simulating several agents interacting in a shared environment is still being researched	Multiple players and NPCs share the same world
Legible text is often only generated when it is in the world description	Showing score, inventory, and menu text
Continuous interaction is limited to a few minutes	Playing a single session for longer than a few minutes
Real-time audio accuracy is still being improved	Matching sound effects and dialogue to events on screen

DeepMind says the drift in object permanence is more noticeable when a simulation runs longer, after several complex prompt edits, or when navigating through the scene. Every feature in the right-hand column is essential to many finished games.

Godot introduction to physics / Genie 3 announcement / Genie 4 model page

In game development, Genie 4 fits the stage before you write engine code

DeepMind has been working with trusted testers in creative industries, journalism, education, gaming, and engineering. Game developers and others are testing whether they can go from an initial idea to an explorable prototype in seconds. They try out visual styles, material looks, and interaction mechanics with text prompts before writing a single line of engine code, DeepMind says. Filmmakers step into a live scene and run multiple takes on the spot, changing camera angles, lighting, dialogue, and action.

According to TechCrunch, Diego Rivas, one of the authors of the Genie 4 announcement, said the real-time aspect is most valuable at the ideation and storyboarding stages. Rivas also said that users could eventually turn such prototypes into creative assets with Gemini Omni, though those capabilities are still some way off.

All of these uses come before the work of building the game in an engine. DeepMind's description includes no example of using Genie 4 as the runtime for a finished game. DeepMind itself names hybrid world models as a direction for future work. The idea is to combine the logical, structured reasoning of code and language models with pixel-level simulation. According to TechCrunch, Tejas Kulkarni, a research director at Google, described a setup in which Gemini writes code and text and Genie takes them in in real time.

Even in DeepMind's vision, a layer that handles rules and structure remains. Whether that layer will be an existing game engine or code written by Gemini has not been shown. This announcement does not answer whether game engines will become unnecessary.

In an article on October 17, LATENT wrote that handing game development to AI requires turning requirements into state changes and checks. On an engine where state can be read out as numbers, those checks can run repeatedly as tests. For now, we think a practical split is to use Genie 4 prototypes to check look and feel, and to build the rules and state on the engine side.

Once Genie 4 is available in Project Genie, LATENT suggests three things to check when judging whether it works for prototyping. First, whether objects and layout are still there when you return to the same place. Second, whether the art style and characters hold up after several text instructions. Third, whether text you want players to read, such as signs or on-screen text, is legible. Each of these corresponds to a limitation DeepMind has listed.

Google DeepMind announcement / TechCrunch article / Related: DeepSeek's new model demonstrates game development in Godot

When and on what terms it becomes available

Genie 4 is currently offered as a research preview for trusted testers. DeepMind plans to expand access starting with Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio. A waitlist for Genie 4 in Project Genie has opened. TechCrunch reported that the rollout will come in the coming weeks. API pricing has not been announced.

Project Genie launched on January 29 for Google AI Ultra subscribers in the US aged 18 and over. On May 19, Google announced that it would gradually roll Project Genie out to eligible Google AI Ultra subscribers globally. That announcement did not list countries or regions. The Genie 4 announcement also does not say in which regions it will become available.

All video and audio generated by Genie 4 will include SynthID watermarks. DeepMind says Genie 4 has safety guardrails consistent with the policies for its other models, and that it went through internal safety evaluations before being offered to testers.

Google DeepMind announcement / Project Genie launch / Project Genie expansion

Sources and scope

Specifications, availability terms, and limitations were checked against primary sources on October 28, 2026. The announcement date follows the date on DeepMind's blog. Statements about Genie 4's performance come from DeepMind's descriptions and published videos, and LATENT has not used the model. Resolution, frame rate, API pricing, and a launch date have not been published. We also could not confirm in DeepMind's materials the two-minute figure reported by TechCrunch. The comparison with game engines, the mapping of limitations to game features, and the suggested checks are LATENT's analysis and suggestions.