Genie 4 is now live in Project Genie for Google AI Ultra subscribers. You can edit a world mid-session by typing, sound is generated with the visuals, and a Gemini agent helps build and remix worlds.
- Genie 4 is now rolling out in Project Genie to eligible Google AI Ultra subscribers.
- You can edit worlds mid-session by typing prompts, for example to change the weather, add characters or redirect the action.
- It's the first Genie model with native audio: ambient sound, music and characters who talk back in real time.
- Gemini runs in the loop as a planner, breaking goals into steps and checking frames to keep worlds coherent.
- Genie 3 now supports 3-minute sessions, world chaining and 360-degree panorama seeding; a new community gallery lets users remix worlds.
- Known limits are drifting details, weak object permanence and no published benchmarks. API access is still pending .
Google DeepMind has moved Genie 4 from private demonstrations into limited subscriber access through the Project Genie browser app. The Project Genie rollout gives eligible Google AI Ultra subscribers four additions: Genie 4, upgrades to Genie 3, an agent for creating worlds, and a gallery for publishing and remixing them.
A world model synthesizes an environment one frame at a time as a person navigates it. Inputs alter subsequent frames, allowing the scene to react during a session. Genie 4 adds mid-session text steering and model-generated audio, capabilities aimed at rapid prototyping for games, film, interactive media, and simulation research.
What shipped, and where #
| Genie availability and capabilities at launch | |
|---|---|
| Area | Launch status |
| --- | --- |
| Access | Eligible Google AI Ultra accounts through the hosted Project Genie app |
| Core interaction | Keyboard navigation and text prompts that modify a running world |
| Audio | Ambient sound and character speech generated with the visual frames |
| API | No public access; early partner access is planned |
| Known limits | Continuity drift, inconsistent physics, audio errors, illegible text, and spatial mistakes |
Prompts rewrite a running world #
During a Genie 4 session, a text instruction can alter the active environment without restarting the generation. Users navigate with standard keyboard controls and can prompt changes to:
- Weather, lighting, and time of day
- Camera position and perspective
- Characters and objects in the scene
- The action unfolding around the user
Genie 4 also generates ambient sound and character speech in real time. Wind, waves, dialogue, and reactions emerge with the scene. In Google’s description, “native” audio means the same model produces sound and visual frames together.
DeepMind reports improvements in aerodynamics, collisions, character movement, and responses to environmental changes. Google has not published benchmark results that quantify those gains, so the claims currently rest on demonstrations and early hands-on use.
Gemini plans between frames #
Google describes Genie 4 as part of a multi-model system. Gemini Omni anchors the simulation to recognizable real-world appearance and behavior, providing guidance about how objects, people, and actions should look. Other Gemini models handle planning and continuity.
Gemini 3.8 Flash serves as a guiding agent inside the generation loop. It observes incoming frames, breaks a goal into steps, tracks progress, and checks whether each step succeeded. When a user submits an edit, Gemini adapts the instruction to the scene’s current context before passing it to Genie 4.
The following pseudocode illustrates the pattern. It does not represent a public Google API:
plan = gemini.plan(goal)
for frame in genie.stream(world):
if user_input:
instruction = gemini.adapt(user_input, frame)
genie.steer(instruction)
if gemini.step_done(plan.current, frame):
plan.advance()
genie.steer(plan.current)
DeepMind calls this approach a hybrid world model. The language model supplies structured reasoning, sequencing, and review. The world model generates the audiovisual environment and responds to steering instructions at the pixel level.
Google has not released a technical paper, full architecture, parameter count, training description, or model card for Genie 4. Those omissions limit independent analysis of how the planner and generator interact.
Project Genie becomes a workflow #
The hosted Project Genie app packages several Google models into a creation and sharing workflow:
- Creation agent: A natural-language description becomes initial visuals and a custom background score through Google’s multimodal models.
- Gemini copilot: Gemini watches the session, suggests possible actions, and can trigger events inside the world.
- Community gallery: Users can publish worlds and remix their structure, soundtrack, or rules.
- Genie 3 upgrades: Generations can run for three minutes, up from 60 seconds. Users can also chain worlds and begin with a 360-degree panorama.
| Choosing between the available models | ||
|---|---|---|
| Model | Best fit | Distinct capabilities |
| --- | --- | --- |
| Genie 4 | Short interactive concept tests | Higher visual fidelity, improved physics, native audio, and mid-session prompting |
| Genie 3 | Longer or panorama-seeded explorations | Three-minute generations, world chaining, and 360-degree image starts |
Access stays gated #
Genie 4 is available to eligible Google AI Ultra subscribers. Google AI Pro members can join the waitlist through the same plan page, and Google does not sell Genie 4 separately.
At publication, Google listed a $99.99 monthly Ultra tier and a $200 monthly option that included Project Genie. Pricing, taxes, promotions, and eligibility can vary by market.
Developers cannot integrate Genie 4 into applications today. DeepMind says it plans to offer access to early partners through the Gemini API and Google AI Studio, but it has not announced a release date, API schema, quotas, or pricing.
Google screens model inputs and outputs for safety. Generated output also carries SynthID watermarks, which identify media created with Google’s generative systems.
Continuity weakens over time #
DeepMind’s research overview lists persistent weaknesses in visual detail, complex physics, audio accuracy, legible text, and spatial consistency. Object permanence, the ability to keep an object stable after it leaves view, can deteriorate during long sessions, repeated edits, or extensive navigation.
Early hands-on reports describe scenes changing after the user turns away and looks back. One account attributed the behavior to on-demand frame generation without a stable, user-visible 3D environment preserving every object. Reviewers also reported perceptible motion lag, despite faster control response than earlier world models.
Google has not disclosed output resolution, frame rate, end-to-end latency, Genie 4’s maximum session length, standardized test scores, or a quantitative comparison with Genie 3. Those missing measurements make it difficult to assess suitability for latency-sensitive applications, repeatable experiments, or physical simulation.
Rapid prototypes fit best #
DeepMind’s trusted-tester program found the strongest uses in workflows where speed matters more than precision:
- Film previsualization: Directors can test camera positions, lighting, dialogue, and action inside a live scene before production.
- Game concept design: Teams can explore visual styles and interaction ideas before building engine assets or gameplay systems.
- Journalism and historical reconstruction: A photograph can seed a walkable interpretation of a place. Such scenes require clear labeling and verification because the model generates unobserved details.
- Robotics and agent simulation: A Google demonstration pairs Gemini 4 Argon with Genie 4 to simulate Apptronik’s Apollo humanoid playing checkers through a sequence of planned actions.
Google says Waymo previously used Genie 3 to build the Waymo World Model, which simulates unusual situations that robotaxis may encounter. The work shows how generated environments can support research beyond entertainment, provided teams validate physical accuracy and scenario coverage.
Competing systems are developing along similar lines. Runway offers its GWM Worlds 2 research preview, and Skywork distributes Matrix-Game 3.0 as open weights under the MIT license. Genie 4’s distinguishing combination is real-time audio, interactive steering, and a language-model planner inside a consumer-facing application.
The planner-renderer split #
Genie 4 currently fits concept exploration and approximate simulation. Its continuity and physics limits exclude production tasks that require stable state, repeatable behavior, or validated physical results.
The reusable design idea is the division of labor between models. A language model decomposes goals, reviews the current state, and adjusts instructions. A world model renders the resulting sights, sounds, and interactions. That architecture could support future agent and robot simulators once developers can measure latency, continuity, and physical accuracy through documented interfaces and reproducible tests.