Google made Genie 4 available through the Gemini API’s Interactions API and Google AI Studio on November 19, 2026. Developers can start with text, a reference image or both, then navigate generated environments and change them while they run. Google describes the output as continuous 720p video at 24 frames per second, accompanied by native, synchronized audio. (Google’s developer announcement)
This is a separate release from the November 18 Project Genie rollout for eligible Google AI Ultra subscribers. That launch introduced Genie 4 through Google’s consumer-facing world-building experience. The November 19 announcement adds a programmatic route and an AI Studio workspace, including an interactive canvas and starter apps.
The distinction matters because Genie 4 is a preview world-generation model, not a conventional text model with a new name. Developers are working with an ongoing audiovisual environment that responds to actions. That creates different application-design problems, especially when a scene must preserve objects, geometry or physical behavior over time.
What Changed With the Developer Release
The November 18 Project Genie announcement covered a broader product update: Genie 4 access for eligible Ultra subscribers, upgrades to Genie 3, a conversational creation agent and a community gallery.
The next day’s announcement names two developer entry points:
- Gemini API, through the Interactions API: The programmatic access route for applications using Genie 4.
- Google AI Studio: A place to explore the model through an interactive canvas and starter apps.
These are related releases, but their access conditions should not be treated as interchangeable. The developer announcement does not say that an Ultra subscription includes API usage, nor does it provide a complete account-eligibility or regional-availability breakdown. Google directs developers to current Gemini API documentation and pricing.
For an initial evaluation, I’d start with the AI Studio canvas or a starter app. The reason is practical: a world model’s value depends on how navigation, live prompts and audiovisual output work together. Seeing that interaction loop is a more useful first step than designing an integration around an assumed chatbot-style request and response.
The opened announcement does not include an exact model identifier, SDK example or request schema. Those implementation details need to come from the current API documentation; an invented sample call would offer false precision.
Genie 4 Generates an Environment, Not a Text Answer
Google DeepMind’s model card, published November 18, describes Genie 4 as a transformer-based system with native support for text, image and user-action inputs. Its output is real-time video with synchronized audio.
The model is built on Gemini Omni Flash and uses Gemini for reasoning and real-time steering. Google describes Gemini Omni as helping anchor simulations in real-world appearance and behavior, while Gemini supplies reasoning intended to support coherence and continuity.
That architecture explains the Gemini connection without turning Genie 4 into a general-purpose text assistant. Text is one way to specify or change the environment. Images provide visual input. User actions influence what happens as the generated scene unfolds.
| Component | What Google documents | Practical significance |
|---|---|---|
| Starting inputs | Text prompts, reference images or both | Developers can describe a setting, supply a visual reference or combine the two. |
| Navigation | Real-time user actions; the announcement highlights keyboard controls | Interaction continues after the initial generation prompt. |
| Live edits | Text prompts affecting weather, lighting, objects or events | Applications can let users reshape a running scene. |
| Visual output | 720p video at up to 24 frames per second | The published output is an audiovisual stream, not an exported 3D scene. |
| Audio output | Synchronized native audio, including ambience, effects and dialogue | Sound can be generated alongside the changing environment. |
The announcement describes 24fps generation; the model card qualifies this as up to 24 frames per second. That is a stated output capability, not an independently measured performance guarantee. It also says little about end-to-end responsiveness: frame rate and the delay between pressing a key and seeing a response are different measurements.
The output specification matters for application planning. Google documents video and audio here, with no geometry-export interface described in these sources. A spatial-design team should therefore distinguish an explorable visual concept from an editable scene containing dependable meshes, dimensions and object properties.
Continuous Generation Does Not Mean Unlimited Consistency
Google says Gemini reasoning helps keep Genie 4’s world state coherent. The model card puts an important boundary around that claim: it says “visual details and object permanence can still shift or drift over time.” (Google DeepMind model card)
Object permanence means that something should continue to exist, with the same identity and relevant properties, when it leaves the camera’s view. In an interactive world, that could mean returning to a room and finding the same furniture in the same places.
Google identifies several conditions that can make drift more likely:
- Longer-running simulations.
- Multiple complex prompt edits.
- Navigation throughout a scene.
It also lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as ongoing research areas.
Those limits have direct consequences. An environment may look convincing while failing to preserve a layout accurately. Audio can be generated in sync with a scene while particular sounds remain imprecise. A sign may be visually plausible without containing reliably readable text.
My view is that persistence is the key evaluation target for developers. Attractive opening frames demonstrate only part of an interactive system’s usefulness. If an application depends on revisiting locations or tracking objects, the relevant question is whether those details survive navigation and edits.
A useful evaluation would therefore include leaving and returning to a location, changing one scene property at a time, and checking whether unrelated details remain stable. These are suggested checks derived from Google’s published limitations, not results from testing Genie 4.
Genie 3’s Three-Minute Sessions Are a Separate Feature
The November 18 announcement says Genie 3 gained sessions of up to three minutes per generation, increased from 60 seconds. It also describes linked worlds and 360-degree panorama grounding for Genie 3. Google explicitly positions that model as an option for longer runs or panorama-seeded spaces. (Project Genie announcement)
Those details should not migrate into Genie 4 specifications simply because both models appeared in the same product announcement.
The November 19 API post describes continuous Genie 4 world generation, but it does not supply a maximum session duration. “Continuous” describes the ongoing generation experience; it is not evidence of unlimited runtime or a three-minute API allowance.
This is a consequential distinction for developers planning a guided experience, training episode or demonstration. Session limits need to be confirmed for the specific model and access route.
Where the Documented Use Cases Make Sense
Google names interactive prototyping, spatial concepts, robotics and AI-agent training as potential uses for Genie 4. These are proposed applications, not evidence that the preview has met a particular production or scientific standard. (Developer announcement)
The strongest fit depends on what an application needs to trust.
Interactive Prototypes
A developer could explore a setting, change the lighting or weather, introduce an object and assess how the experience changes. Native audio adds another part of the concept, from ambient sound to character dialogue.
Short, exploratory prototypes align well with a model that Google says can maintain coherent environments over short sessions. The tradeoff is reproducibility. If an exact object arrangement or event sequence is essential, the prototype needs explicit checks for those properties.
Frequently Asked Questions
3 questions
1Can Genie 4 Start From a Reference Image?
Yes. Google says the developer release accepts text, a reference image or both. Users can then navigate and modify the generated environment. (Developer announcement)