Genie 4 adds sound and puts Gemini in the director's chair. Google skipped the performance numbers

Google DeepMind has announced Genie 4, a world model you can step into, move around with a keyboard and redirect with text prompts. What's new is native audio, including dialogue, and a Gemini model that supervises the simulation as it runs. What's missing are the hard numbers Google still gave for its predecessor.

Genie 4 generates worlds with sound and lets Gemini keep watch

Google DeepMind announced Genie 4 on October 27, 2026. Users move through a generated environment with the keyboard, control the camera and change the scene on the fly with a prompt: add rain, a new character or an object. DeepMind says the model is natively multimodal, so music, ambient sound and dialogue from characters who react to the player are generated together with the world. The company calls it its first real-time world model to generate audio natively.

The interesting part is what surrounds the model. Gemini Omni keeps the world's look and behaviour close to reality. Gemini 3.8 Flash runs in the loop as a guiding agent: it watches the scene, breaks a goal into steps and checks incoming frames to verify each step actually happened. In a robotics demo, Gemini 4 Argon plans a task for Apptronik's Apollo humanoid playing checkers, and Genie 4 renders each step.

For now, only trusted testers have access as a research preview. Next in line are Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio. Everyone else can join a waitlist. All Genie 4 video and audio is meant to carry a SynthID watermark.

Google sells it as play, but the targets are robot training and film previs

The comparison point is August 2025. For Genie 3, DeepMind gave concrete specs: 720p, 24 frames per second and consistency for a few minutes. Project Genie, the public interface built on Genie 3, opened in January 2026 only to AI Ultra subscribers in the US aged 18 and over. Genie 4 shifts the emphasis from looking at a world to simulating events: you set something in motion and the model works out how the world responds and what it sounds like.

The section on testers shows who the real customers are. Filmmakers run multiple takes inside a simulation, changing camera, lighting and dialogue. Game developers get a prototype before writing a line of engine code. Photojournalists are testing how to turn an archival photo into a space you can walk through. Enterprises want fast scenario generation for simulation work.

Robotics is the most consequential direction. Instead of hand-built environments, the model can produce endless variations of a kitchen or a warehouse for a robot to learn in. DeepMind closes by describing hybrid world models: a language model plans and keeps the logic of events, a pixel model draws the physical reality. Anyone building agents will recognise the familiar split between planner and executor.

Without numbers on duration and physics, realism stays a promise

This is where the announcement is weak. DeepMind gave no resolution, frame rate, latency or figure for how long a world stays coherent. It admits that details and object permanence drift over longer simulations, especially after several complex prompt edits or while moving through the scene. Complex physical interactions, audio accuracy, legible text and spatial accuracy remain open research problems, by the company's own account.

For robot training, that matters. A robot learning to pick up a mug in a world where the mug quietly turns into a different mug is learning from fake data. The blog's clips are curated recordings driven by long, carefully tuned prompts, not a random sample. SynthID reduces the risk of fabricated footage, but only for people who can and want to check the watermark.

The Gemini API price and the first robot trained in Genie will decide it

The key signal will come with the Gemini API. Once developers get Genie 4, we will learn the price per minute of simulation, the latency and the real coherence window, exactly what the blog left out. The second signal is robotics: if Apptronik or another partner shows behaviour trained in Genie 4 that works on real hardware, it will be a stronger argument than every saxophonist floating on a lake.

For European readers, availability is an open question. Project Genie launched in the US only, and Google has not named any countries for Genie 4 yet.

Lilith's verdict

Google gave the saxophonist on the lake a voice and sat Gemini next to him with a checklist. An engineer who wants to teach a robot to pick up a mug in that world will first ask the one thing the blog left out: how long does the mug stay the same mug.