### The new part is the soundtrack, generated live
Earlier Genie demonstrations focused on explorable visuals. Genie 4 adds a soundtrack generated as part of the simulation itself. DeepMind calls Genie 4 its first real-time world model to generate audio natively. DeepMind says the audio is generated alongside the world rather than added as a soundtrack afterwards. Its examples include ambient sounds such as wind in trees and waves at a beach, along with character dialogue that responds to events in the scene and the user's actions.
Native audio is the more consequential change in this announcement. The demos offer a look at the visuals, but without published performance figures or comparable tests, they cannot establish how much the underlying visual model has improved. A world that talks back changes what the simulation can be used for. A scene you can rehearse, not just look at, is closer to a rehearsal space for a film, a lesson or a robot task than to a video clip with controls attached.
Steering has changed in the same direction. A user can type an instruction while the simulation runs, change the weather, move the camera, add a person, and Gemini is meant to fit that instruction to the scene rather than drop it in raw. Between user inputs, Gemini monitors the scene and keeps it moving in a coherent direction.
### Two models, one of them checking the other
Genie 4 does not work alone in DeepMind's account. Gemini Omni anchors the simulation in real-world appearance and behavior. Gemini 3.8 Flash sits in the loop as a guiding agent: it watches the frames Genie produces, breaks a goal into a sequence of steps, checks whether each step succeeded, and rewrites its plan when the user intervenes.
DeepMind's robotics example shows where this is aimed. In an experiment with Gemini Robotics, a separate model, Gemini 4 Argon, broke a task into steps and fed them to Genie 4 as captions while it rendered, simulating an Apptronik Apollo humanoid playing checkers. The experiment demonstrates how a separate planning model can direct a generated simulation step by step. It does not establish that the environment is accurate or consistent enough for robot training, but it shows the kind of interactive simulation workflow DeepMind is pursuing.
### What the announcement does not give you
DeepMind's announcement describes what Genie 4 can do but gives no resolution, frame rate, numerical session limit, benchmark results or model size. It also does not link to a technical paper. Those omissions make it difficult to compare the model with its predecessor or assess how its demonstrations translate into repeatable performance.
DeepMind describes Genie 4 as a leap in photorealism and physical realism, including more natural human movement and improved simulation of collisions and aerodynamics. Those are the company's qualitative claims, illustrated through selected demonstrations rather than supported by published comparative benchmarks.
The one number now circulating, that Genie 4 can hold a consistent scene for up to two minutes, twice its predecessor's limit, comes from press briefing reporting by TechCrunch and Bloomberg. It does not appear in DeepMind's published announcement. DeepMind's own wording is vaguer and less flattering: the model maintains coherent environments "over short sessions."
Its limitations section deserves more weight than its adjectives. A research preview need not arrive with production-ready specifications. But without comparable measurements, the announcement cannot establish how much of the claimed progress carries over to ordinary use. DeepMind says visual details and object permanence can shift or drift over time, particularly on longer runs, after several complex prompt edits, or while navigating. It lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as active research problems. Audio accuracy is on that list in the same announcement that leads with audio. Both halves are the company's own words.
### Who can use it, and for what
For now, trusted testers. DeepMind says filmmakers in that group are using Genie 4 as a draft mode, stepping into a scene and running several takes while adjusting camera, lighting, dialogue and action. Photojournalists are testing whether one archival photograph can become a simulation of the place. Game developers are using it to test an idea as an explorable prototype before writing engine code.
Everyone else gets a waitlist. DeepMind says Genie 4 is coming to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, and gives no date in its own post. Until that rollout happens and someone outside the tester group can measure a session, the practical product here is a set of demo recordings and a direction of travel.
Sources
- [1] Google DeepMind, "Genie 4: A new generation of world simulation," October 27, 2026Read source
- [2] Google, "We're announcing Genie 4, our new world model for real-time interactive simulation," blog.google, October 27, 2026Read source
- [3] TechCrunch, "Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering," October 27, 2026Read source