Genie 4: Google's Interactive World Model Now Generates Sound

Sources checked and dated on publication. See our editorial policy.

On October 27, 2026, Google DeepMind announced Genie 4, a research-preview world model that generates navigable environments with ambient sound and interactive dialogue, and lets you change a running scene with plain-language prompts. It is not a product you can buy today: access is limited to trusted testers, with a waitlist for Project Genie. Google has published no price, no API model ID and no general-release date.

What a world model actually is

Related guide: Anthropic's red team says AI worms could soon spread with no datacenter — here's what their tests actually showed

Interestingly, neither Genie 4 post defines the term. Google's own definition comes from its August 2025 Genie 3 post: world models are "AI systems that can use their understanding of the world to simulate aspects of it, enabling agents to predict both how an environment will evolve and how their actions will affect it." That is a different job from a chatbot or a video generator: the system is expected to hold a scene together while someone acts inside it.

Genie 4's demos are, per DeepMind, "recordings of real-time interactions" in which users navigate with standard keyboard controls and steer the scene with text prompts as it evolves. Google frames the point as simulation rather than content generation. Its short post calls simulating the world "a crucial step toward AGI," citing safe training environments for robots and agents and creative tools in filmmaking, journalism, gaming and manufacturing. The longer post describes trusted-tester cohorts in creative industries, journalism, education, gaming and engineering, testing uses such as film pre-visualization, turning a single archival photograph into a grounded simulation, and prototyping a game scene "before writing a single line of engine code." These are Google's stated ambitions and tester use cases, not independently evaluated results.

Native audio is the headline change

Related guide: Google's Gemini Model Gate: What Free and AI Plus Users Lose From October 9

DeepMind calls Genie 4 "our first real-time world model to generate audio natively." Every world carries ambient sound of its own setting — the post gives wind in the trees and waves at the beach as examples — and characters can hold interactive dialogue that "adapts on the fly," reacting to your inputs or to events in the world in real time. Google's shorter post adds music to that list of generated audio: "ambient sounds, music, and interactive dialogue."

There is a caveat worth keeping in the same paragraph. The detailed DeepMind post lists "real-time audio accuracy" among its active areas of ongoing research, and TechCrunch reported that the team had not yet worked out how to guarantee the audio is accurate to the scene. Native audio therefore means sound is produced as part of the world, not that it is reliably faithful. Google also says all video and audio generated by Genie 4 will carry SynthID, its imperceptible watermarking for identifying AI-generated content across visual and audio modalities — a provenance measure, not a quality metric.

Steering a running scene, with Gemini behind it

Related guide: Gemini 4 Argon is the most powerful model Google has built — and you cannot use it yet

The interaction model is the second substantive change. Google says you can "use natural language to modify a simulation in real-time — like changing the weather, moving the camera, or introducing a new person or object — and Gemini will contextually adapt your prompt to match the surrounding scene." Between inputs, Gemini "monitors and steers the scene to keep the world progressing cohesively."

The longer post names the models, and the names matter because they are easy to misreport. Gemini Omni "anchors the simulations in real-world appearance and behavior." Gemini 3.8 Flash serves as "a guiding agent in the loop": it observes the scene, breaks goals into steps and reviews incoming frames to check whether each step succeeded, adapting its plan when the user intervenes. The Verge describes this as Gemini 3.8 Flash working "under the hood." Separately, Google says it has been experimenting with Gemini 4 Argon alongside Genie 4 for Gemini Robotics, where Argon breaks a user-instructed task into steps that are fed back to Genie as text captions while the simulation renders. That is described as an experiment, not a shipped product feature. Google's short post names only Gemini Omni; the other two names appear in the DeepMind technical post.

On physics and realism, the DeepMind post says Genie 4 simulates scenarios such as aerodynamics and collisions, and that "the most dramatic leap" is how it simulates people, with characters that move naturally and respond to their environments. Both Google posts use the phrase "unprecedented realism." These are Google's characterizations and tester impressions; no published benchmark, resolution, frame rate or independent replication accompanies them. One honest comparison point comes from Google's own Genie 3 post, which specified 24 frames per second at 720p and consistency over "a few minutes" for the earlier research model — figures that do not appear in either Genie 4 post.

Access, pricing and the two-minute number

Related guide: The UK can test frontier AI models again: what the 17-day freeze and the U-turn actually changed

Genie 4 is "a research preview for trusted testers," with access "coming soon" to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio; Google links a waitlist form. heise online states plainly that Google has not announced a date for general release. The Verge's reporter notes they had not been able to try it. TechCrunch says the wider rollout is expected "in the coming weeks," which is a report of expectation, not a dated commitment. No Genie 4 price, per-token rate, API model ID or rate limit appears in any primary post.

One widely repeated number deserves careful handling. TechCrunch and The Verge report that Genie 4 can hold a scene for about two minutes, up from one minute for the previous version; heise attributes the two-minute figure to Bloomberg. Google's own Genie 4 posts do not give a duration — they say only that Genie 4 "can maintain coherent environments over short sessions." Two minutes is therefore press-briefing reporting, not an official specification. It also sits awkwardly beside Google's Genie 3 post, which claimed consistency over "a few minutes" for the research model, and beside reporting on the earlier consumer-facing Project Genie, which had a roughly 60-second session cap. The likely explanation is that different things are being measured — a product session limit versus a research model's coherence — but Google has not published a comparison table that would settle it.

More useful than the ceiling is the qualification Google does publish. Its limitations section says that while coherent over short sessions, "visual details and object permanence can still shift or drift over time," especially as a simulation runs longer, after multiple complex prompt edits, or when navigating around the scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are all listed as "active areas of ongoing research." Google calls Genie 4 "an early, experimental model." Reaching a two-minute session, in other words, would not imply two minutes of stable world state.

Safety framing and what to watch next

Related guide: Trojanized Claude Code Extension: What the Cursor Backdoor Teaches

DeepMind says Genie 4 "includes strict safety guardrails" consistent with its policies across other models, that it ran internal safety evaluations before the preview to prevent harmful or malicious content, and that it intentionally deployed first to trusted testers for feedback before a broader release. It does not publish an evaluation methodology, a red-team report or a system card in these posts. The SynthID watermark covers both video and audio output. Readers should treat the safety claims as stated policy, not as independently audited evidence.

For developers, the practical position is straightforward. There is nothing to integrate yet: no public API, no model ID, no price, no SDK instructions in the launch posts. The Gemini API and AI Studio are named as future early-access routes, which is a statement about planned distribution, not about Genie 4 being a Gemini LLM or about Gemini token pricing applying to it. Anyone planning work around generated interactive worlds should wait for an actual release announcement with published limits and terms, and should not budget against the unconfirmed two-minute figure. Google's own framing — "towards hybrid world models," combining the structured reasoning of code and language models with pixel-level simulation — describes a research direction more than a finished product.

Bottom line: Genie 4 is Google's first real-time world model to generate audio natively, steerable mid-simulation with plain language and kept coherent in part by Gemini models in the loop. It is a limited research preview with no published price, specs or release date. The careful reading is to credit the capability, discount the unverified numbers, and wait for documentation before planning anything on top of it.

Prices and benchmark figures are standard published rates as of November 5, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.