news

Google's Genie 4 arrives with native audio, trusted-tester access and no published numbers

DeepMind's world model generates sound and dialogue alongside the scene and leans on Gemini Omni for appearance, but the only duration figure in circulation came from a press briefing.

By Daniel Ochoa · 28 October 2026

Google DeepMind announced Genie 4 on October 27 as a real-time interactive world model, and the announcement carries no resolution, no frame rate, no benchmark and no stated session length. What it carries is a capability list and a rollout that starts with a limited group of trusted testers.

Audio generated with the scene, and Gemini models inside the loop

DeepMind describes Genie 4 as natively multimodal, with music, ambient noise and interactive dialogue produced at the same time as the world itself, which it calls a first for a real-time world model in its line. Users steer scenes as they run using plain language, changing weather, moving the camera or introducing new characters, and the characters can speak and react to those inputs. The lab says the model "unlocks unprecedented realism," a phrase doing a lot of work in a post that offers no measurement to attach it to.

The plumbing is explicitly Gemini. DeepMind says Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a simulation coherent, with Gemini 3.8 Flash acting as a guiding agent that watches the scene and plans multi-step events as they unfold. All generated video and audio is watermarked with SynthID.

The two-minute figure came from a briefing, not the post

The number most widely attached to Genie 4 is a roughly two-minute window of scene consistency, double what Genie 3 managed, along with improved physics and more humanlike characters. That came out of a press briefing reported by TechCrunch rather than from a published spec sheet. For an interactive simulator, how long a world holds together before it drifts is close to the whole product, and it currently lives in press coverage instead of documentation.

DeepMind is open about the drift itself. Its own limitations note that visual detail and object permanence can shift over longer runs, after several complex prompt edits, or while a user navigates, and legible text in a world tends to appear only when it was written into the prompt. Those are useful admissions, and they are also the closest thing the launch offers to a failure-rate disclosure.

A pre-visualization tool competing with storyboards, not finished clips

Filmmaking sits first in DeepMind's list of industries the model could serve, ahead of journalism, gaming and manufacturing, and testers are reported to be using it for VFX pre-visualization. Google DeepMind engineer Diego Rivas framed the value of real-time generation around rapid ideation, concept development and the storyboard stage, and said repeatedly during the briefing that this is early. That positions Genie 4 against the planning phase of a production rather than against a model that hands back a finished shot, which is a different market from the one Veo 3.1 and its successors sell into.

The contrast with the model underneath it is sharp. Gemini Omni, the thing DeepMind says anchors Genie 4's appearance and behavior, is already a purchasable generation endpoint across several platforms on our price radar, while Genie 4 has no price, no date and no country list.

Access starts narrow

Genie 4 is rolling out to a limited group of trusted testers now, with Google AI Ultra subscribers getting it through Project Genie and early access partners reaching it through the Gemini API and AI Studio at an unspecified later point. A waitlist is open. Until outside users can run it, every claim in the launch, including the audio-first claim, rests on curated demos chosen by the lab that built the model.

Sources: Google DeepMind, Google, TechCrunch.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.