Google DeepMind introduced Genie 4 on October 27, 2026, describing it as a world model that generates interactive environments in real time, with generated audio and physics. It is a research preview for trusted testers, with wider access planned through Project Genie for Google AI Ultra subscribers. Everything below comes from DeepMind's announcement unless another source is named.
GoogleWhat Genie 4 is
A world model generates a simulated environment frame by frame in response to a user's inputs, rather than playing back a fixed video. DeepMind says Genie 4 is natively multimodal, producing music, ambient noise and interactive dialogue along with the picture. Users navigate and control the camera with standard keyboard controls, and can steer the scene in natural language, for example by changing the weather or adding a person or object.
DeepMind also says the model simulates real-world physics such as aerodynamics and collisions, and that characters in the world move and speak in response to events. The announcement links back to Genie 3, published in August 2025, but does not give a side-by-side specification for the two.
Natively multimodal, enabling audio-like music, ambient noises, and interactive dialogue.
Google DeepMind, describing Genie 4
Gemini as the guiding agent
The most structurally interesting detail is that Gemini sits beside the world model as an agent. According to DeepMind, Gemini adapts user prompts to fit the scene, and acts as a guiding agent that observes frames, breaks goals into steps, verifies each step and adapts when the world changes or a user intervenes. The page names Gemini 3.8 Flash for this role.
The Genie 4 and Gemini guiding-agent loop
As described by Google DeepMind. The loop repeats for as long as the session runs.
DeepMind also describes experiments pairing Genie 4 with a Gemini 4 Argon model and with Gemini Robotics, including a simulated checkers game with Apptronik's Apollo humanoid. Our earlier coverage of Gemini 4 Argon explains that model's launch.
What the announcement does not say
The announcement leaves out several numbers a technical reader would look for. It does not state output resolution, frame rate, or how long a session can run. It does not publish benchmarks against Genie 3 or other world models. Those gaps matter because a world model's usefulness depends on them: a low frame rate or short session limits previsualization, game prototyping and agent training alike.
DeepMind is open about known weaknesses. It says visual details and object permanence can drift, especially in longer sessions, after multiple prompt edits, or while moving through a scene. It lists complex physical interactions, real-time audio accuracy, legible text rendering, spatial accuracy and long-term consistency as ongoing research areas.
Why consistency is the real test
Object permanence is the property that an item left in a room is still there when the camera returns. For entertainment, drift is a visual flaw. For uses DeepMind lists, such as turning archival photos into simulations or generating scenarios for enterprise simulation, drift undermines the premise, because a simulation that quietly changes its own state is hard to trust. The same applies to robotics: a planner verifying each step against frames only works if those frames stay consistent with each other.
This is an analytical inference, not a DeepMind claim. The company's own list of limitations points the same way: the open problems it names are mostly about staying consistent over time.
Safety and provenance
DeepMind says Genie 4 carries safety guardrails matching its other models, that internal safety evaluations ran before the preview, and that the model went to trusted testers first. It also states that all generated video and audio includes SynthID, its imperceptible digital watermark. Our explainer on Google's SynthID detector covers what that watermark can and cannot establish.
What this means for builders
- World models are becoming a component, not only a demo. Pairing one with a planning model is a pattern that could extend to testing agents in simulated environments.
- Wait for measurements. Until resolution, latency and session limits are published, claims about production use are premature.
- Model choice will fragment. Teams may use different models for text, images, video and simulation. Metir is a model-agnostic workspace where work can move across providers, which matters as specialised models like this appear.
Sources:
- Genie 4: a new generation of world simulation - Google DeepMind
- Everything That Happened in AI Today, Wednesday, October 28 - The Neuron
Image credits
Header and in-body image: Google building at 6 Pancras Square, London, by Gcirian, Wikimedia Commons, CC BY-SA 4.0 (source). It is an older photograph and does not show Genie 4.