Google DeepMind has put Genie 4 out as a research preview, a little over a year after Genie 3. The headline features are easy to list. It generates audio in real time (music, ambient noise, spoken dialogue you can interact with). You can type prompts mid-simulation to change the weather, swing the camera, or drop in new people and objects. Scenes stay consistent for up to two minutes, double Genie 3’s limit, and Google says physics and characters are more convincing.
Audio is the feature that will carry the demo reels. I think the more consequential change is quieter.
Gemini runs the show now
In Genie 4, Gemini acts as a director. Per MarkTechPost, Gemini 3.8 Flash plans events inside the simulation loop, and TechCrunch reports that Gemini plans steps and checks frames while Genie handles the pixels. That is a real architectural shift. Earlier Genie models were impressive dream machines that you poked with a controller. This one has a language model deciding what should happen next and auditing whether it did.
Tejas Kulkarni, a research director at Google, described where this goes: “hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time.” The robotics experiment shows the idea in practice. A Gemini 4 Argon model broke a task into steps, and Genie 4 simulated each one live, a cheap way to rehearse robot behavior before anything moves in the physical world.
I buy this direction completely. Pure pixel prediction drifts, and Google admits Genie 4 still does: details and objects wander the longer a run goes. A planner that checks frames is the obvious patch, and it’s the part of this release most likely to survive into whatever ships.
Two minutes is a storyboard, not a game
Google keeps pitching gaming alongside education and robotics. I don’t see it yet. Two minutes of coherence is a big jump on paper, and it’s still shorter than most tutorials. Runway’s GWM Worlds 2, released a couple of months ago, also offers playable worlds with audio, so neither company can claim the gaming angle as a moat, and neither has shown a persistent world you’d play for an evening.
The trusted testers seem to agree. They’re pointing Genie 4 at film prototyping, and Google DeepMind engineer Diego Rivas leaned into it, calling real-time generation most valuable “in the ideation phase, in the storyboarding phase.” That’s the honest use case. A director can walk through a rough scene, ask for rain, move the camera, add an extra, and hear the room tone, all without leaving the flow. Two minutes is plenty for that. Drift matters less when you’re throwing the output away after the meeting.
Rivas also floated handing those prototypes to Gemini Omni to turn them into finished assets, while conceding that’s “a ways off.” Google already has Veo for polished clips with synced audio. The pipeline it’s sketching runs from Genie for blocking a scene to something else for the final render.
The known gaps line up with that read. Google says it hasn’t figured out how to make the audio reliably accurate or keep in-world text legible. Those are dealbreakers for a shipped game and minor annoyances on a whiteboard.
Rivas said repeatedly that the end-to-end product isn’t defined yet. What nobody at the briefing addressed is how much of Genie 4’s coherence comes from Gemini’s frame-checking, and whether that holds past the two-minute mark.