# Google's Genie 4 adds native audio, dialogue and text steering to its world model > Google DeepMind's Genie 4 generates more photorealistic interactive worlds with native audio and dialogue, runs for about two minutes, and takes text edits mid-simulation. It is a research preview for trusted testers. > > OpenCheese AI · 2026.10.28 · Web page: https://opencheese.dev/blog/google-genie-4-world-model-audio-dialogue-text-steering One of Google's demo prompts for Genie 4 puts you on a black dragon doing 140 km/h, wingbeats included. The same announcement admits that objects in a world like that can drift and change if you stay long enough. That tension runs through the whole launch. Google DeepMind released Genie 4 on Tuesday as a research preview for trusted testers, and it is the most photorealistic and interactive version of the company's world model so far, with sound, speech and edits that arrive while the simulation is still running. ## What Genie 4 changes Genie 3 appeared as a research preview a little over a year ago, and it powers Project Genie, Google's tool for generating interactive worlds from text or image prompts. Genie 4 is the follow-up, and Google describes four main updates. - **More realistic worlds.** Better photorealism, smoother movement, more convincing physics and more human-looking people. - **Edits while it runs.** You type a natural language prompt and the scene changes in real time. Weather, camera position, a new person or object are all fair game. - **Native audio.** Music, ambient noise and spoken dialogue are generated together with the world rather than bolted on afterwards. - **Gemini in the loop.** Gemini 3.8 Flash helps run the simulation, which gets its own section below. Session length also doubles. Genie 4 can hold a consistent scene for up to about two minutes, against roughly one minute for Genie 3. Two minutes is not a feature film, but it is twice what the previous model managed, and in this field that counts as progress. Users move through a world with standard keyboard controls for navigation and camera, and steer it with text prompts as it evolves. Google's own material says Gemini adapts each prompt to fit the surrounding scene, so what you type is a suggestion that gets reworked to suit the world you are standing in. ## Gemini as the director The most interesting change is not visual. Genie 4 builds on the Gemini family, and Google splits the work among several models. Gemini Omni anchors the simulation in real-world appearance and behavior. Other Gemini models supply the reasoning that keeps a simulation coherent and continuous. And Gemini 3.8 Flash sits inside the live loop as what Google calls a guiding agent. While Genie generates the world, Gemini watches the scene, breaks a goal into steps, reviews incoming frames to check that each step actually worked, and revises its plan when the world shifts or the user barges in. Between your inputs, it also steers the scene so things keep moving along sensibly. Google's claim is that this produces worlds with more reliable causality, procedural behavior and multi-step reasoning. The post names Gemini Omni, Gemini 3.8 Flash and, later, Gemini 4 Argon, so the phrase "Genie 4" now covers a fair number of Geminis. Tejas Kulkarni, a research director at Google, told reporters at a briefing where he thinks this goes: "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." Google's framing is ambitious. The announcement calls understanding and simulating the world a core part of building AGI, and pitches Genie 4 as a source of safe, infinitely variable simulations for training robots and agents, as well as new interactive experiences. ## Sound, speech and people DeepMind calls Genie 4 its first real-time world model to generate audio natively. Every world carries the ambient sound of its own setting, from wind in trees to waves on a beach, and all the sounds in Google's demo clips were produced directly by the model. Speech is the part Google treats as the real frontier. Characters can converse and react to your inputs or to events in the world, with dialogue that adapts on the fly. The other leap Google claims is in people. The company says characters now move naturally, respond to their surroundings and can be steered in real time. On the physics side, it says Genie 4 simulates scenarios such as aerodynamics and collisions, built on what it calls an advanced physical understanding of the world. All of this comes from Google's own demos, and at least one journalist covering the launch said they had not been able to try the model. Comparisons with Genie 3 therefore rest on Google's own clips for now. ## What the demo prompts reveal Google published the prompts behind many of its clips, and they are a better guide to the model than the adjectives. Some are short. A crowned crane walks across snow beneath a dusk sky. A macro shot of a maple seed falling into grass. A whole green-striped watermelon on a stone plinth, with airborne droplets backlit in slow motion. Others read like a brief from someone who has been burned by a production company before. The 1980s Japanese train station prompt insists on a dense crowd and a red train on the nearest track, then rules out empty platforms, a distant train, sparse crowds and a symmetrical composition. The dragon-race prompt demands a strict first-person view and then lists, at length, every camera angle it does not want. A saxophone player on a floating platform in a misty lake comes with an order of absolutely no HDR. The range is wide. Other examples include: - a stream of bubblegum-pink satin winding through a wildflower meadow - a first-person hot air balloon ride over Antarctica - a fully stained-glass landscape where even the grass and sky must stay glass - a 1995 Hong Kong romance on a motorcycle at dawn, shot on 35mm - a first-person view of a wizard's wand in a room packed with porcelain animals - a magician beside a table, with a real peacock standing in an open box The point of showing them is that the model follows long, specific, stylistically fussy instructions, and that you can then keep editing the result with more text. ## Testing robots in a world that does not exist The robotics demos show the most distinctive idea in the release. One clip has a Universal Robots UR5 holding a pipette and dispensing liquid into a lab plate, with tiny electric sparks. A more elaborate one pairs Genie 4 with Gemini 4 Argon, in a research experiment tied to Gemini Robotics. The mechanics are simple to describe. Argon takes a task from the user and breaks it into steps. Those steps are fed to Genie 4 as text captions while the simulation is rendering, and since Genie 4 can be steered by text in real time, the simulation can adapt as it progresses toward finishing the task. In the showcase example, Apptronik's Apollo humanoid plays a game of checkers, with the task split into real-time steps. Google's argument is that this is a way to test how AI systems handle different environments before they meet the real one. It is presented as a path rather than a finished result, and no benchmark figures accompany it in what Google released. ## What the trusted testers are doing Google has been working with cohorts of trusted testers in creative industries, journalism, education, gaming and engineering. It reports four kinds of use. - **Creative direction.** Filmmakers and technologists use Genie 4 as a "draft mode" for pre-visualization. They step into a live simulation of a scene and run several takes, changing camera angles, lighting, dialogue and action on the fly. - **Historical scenarios and immersive photographs.** Photojournalists are turning a single archival 2D photograph into a simulation of that place within seconds. Teams are also testing populated, dynamic worlds to explore historical spaces or to visualize short-term events such as a flash flood or wildfire. - **Concept design for interactive experiences.** Game developers go from an idea to an explorable prototype in seconds, trying visual styles, materials and mechanics through prompts before writing any engine code. - **Simulation.** Enterprise partners are evaluating Genie 4 for rapid "zero-to-one" scenario generation, and telling Google where it falls short. The filmmaking angle drew attention at the press briefing. Diego Rivas, a Google DeepMind engineer, said "The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful." He pointed to ideation and storyboarding, or any moment that calls for rapid ideation. Google already sells Veo, its video generation model, which makes high-definition clips with synchronized audio from text, image or video prompts and has been gaining ground among filmmakers. Rivas suggested users could eventually use Gemini Omni, a multimodal model that generates video from images, audio and text, to turn Genie prototypes into finished creative assets. He said that capability is still a way off. An earlier speculative example of interactive learning would let a student enter a painting of a historical scene and chat with the people in it. ## Where it still breaks The limitations section is blunt, and refreshingly specific. Genie 4 can keep an environment coherent over short sessions, but visual details and object permanence can shift or drift. That gets worse as a simulation runs longer, after multiple complex prompt edits, or while moving around the scene. Google lists further areas of ongoing research: - complex physical interactions - real-time audio accuracy - rendering legible text - spatial accuracy - long-term consistency At the briefing, the team added that it has not yet worked out how to guarantee accurate audio or readable in-world text. Rivas stressed that it is early days and that the end-to-end product experience is not yet clear. The watermelon on the plinth, in other words, is safe for two minutes. Beyond that, the model's own authors make no promises. ## Guardrails and watermarks Google says Genie 4 carries strict safety guardrails consistent with its policies for other models, and that it ran internal safety evaluations meant to prevent harmful or malicious content before launching the preview. Limiting the first release to trusted testers is part of the same approach, since it gathers feedback before a broader rollout. All video and audio generated by Genie 4 will include SynthID, Google's imperceptible digital watermark for identifying AI-generated content in both visual and audio form. For a model that generates convincing dialogue alongside photorealistic people, that is a sensible thing to attach. ## Access, and the company it keeps Genie 4 is available now as a research preview to trusted testers. Google says it is coming soon to Google AI Ultra subscribers in Project Genie, and to early access partners of the Gemini API and Google AI Studio. Reporting at launch put the rollout in the coming weeks. A waitlist for Genie 4 in Project Genie is open. The release lands in a busy and consolidating market. AI video startup Runway released its second world model, GWM Worlds 2, a couple of months ago. It is a playable model that reacts to user inputs and generates audio. Meanwhile, Fei-Fei Li's World Labs is being acquired by AMD, Amazon is reportedly in talks to buy Decart, and Rivian has acquired Third Dimension to bring simulation in-house for self-driving testing. Genie 4 is Google's move to stay ahead in that crowd. ## Where this is heading Google's stated direction is hybrid world models. The idea is to pair the structured reasoning of code and language models with the rich, sensory pixel-level simulation of a model like Genie. Gemini maps the structure and plans the narrative timeline behind the scenes, and Genie renders how the world looks, sounds and responds, pixel by pixel. Google argues that this combination should be more coherent, controllable and grounded than either approach alone. Whether it is will depend on the parts the announcement leaves honestly unfinished. A model that can stage a 1995 Hong Kong dawn on 35mm film still cannot promise that a sign in the background will be readable. ## Related posts - 2026.10.06 · [Google DeepMind’s SynthIDBio embeds watermarks in AI-designed protein sequences](https://opencheese.dev/blog/google-deepminds-synthidbio-watermarks-ai-designed-protein-sequences.md): SynthIDBio adds a cryptographic watermark to protein sequences generated with ProteinMPNN, offering a new signal for screening unfamiliar designs. Short proteins and some model types remain outside its reach. - 2026.10.18 · [Google DeepMind's AlphaProof Nexus solves 9 open Erdős problems, two of them 56 years old](https://opencheese.dev/blog/google-deepmind-alphaproof-nexus-solves-erdos-problems.md): Google DeepMind's AlphaProof Nexus agent solved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures, with every proof checked in Lean. The authors are frank about its limits. - 2026.10.18 · [Google's Brin Says Gemini Is Helping Build Itself, but DeepMind's Own Report Sees No Speedup](https://opencheese.dev/blog/google-brin-gemini-self-improvement-deepmind-report-no-speedup.md): Sergey Brin says Google is entering recursive self-improvement with Gemini 4 Argon. DeepMind's safety report, published the same week, finds no meaningful acceleration of AI research and rates the risk as low for now.