Genie 4 adds sound and conversation to Google's AI worlds
Google DeepMind introduces Genie 4, a model capable of generating interactive photorealistic scenes with native audio. The system is released as a research preview for trusted testers, before wider access for AI Ultra subscribers and Gemini API partners.
Google DeepMind has unveiled Genie 4, a world model that generates photorealistic scenes in real time and allows interaction, including sound produced natively by the system. Announced on October 28, 2026, in The Rundown newsletter, it is initially aimed at trusted testers, then AI Ultra subscribers and Gemini API partners. This development is significant for research, video games, and robotics, as it brings simulation closer to an environment where one can not only move around but also talk to the characters inhabiting it.
Des mondes simulés qui répondent, enfin
Genie 4 is the first version in the Genie family to generate its own sound, whether ambient, effects, or music. Characters in these worlds can also hold live conversations with the user, going beyond simple interactive scenery. In a demonstration cited by the source, an Apptronik Apollo humanoid robot plays checkers in one of these environments, with Gemini 4 Argon responsible for planning moves. Altogether, this gives Google a tool that resembles both a simulator and a testing ground for agents and robots.
The technical promise remains clear: the user moves with the keyboard, issues text commands to modify the scene, and the environment maintains coherence between interactions. In practice, this means the world does not reset at every command, which is precisely the small luxury static mockups cannot offer. For studios, researchers, or product teams, this continuity changes how situations, behaviors, and dialogues are prototyped.
Un aperçu de recherche, pas un produit grand public
The model is launched as a research preview, reserved for testers deemed reliable. Google then plans access for its AI Ultra subscribers and Gemini API partners. No general release date is given at this stage [to be verified]. This deployment sequence suggests a cautious strategy: validating uses and limits before broader opening, especially on a system that combines visual generation, audio, and real-time interaction.
The positioning also reveals the direction taken by Google DeepMind. Genie no longer serves only to produce artificial worlds to look at; it tends toward a staging tool, even a preview tool, with obvious applications in video games and robot training. The shift from silent scenery to a speaking world is not trivial. It mainly opens a new category of use and a new ground for comparison with other simulation models.
Ce que Google cherche à tester avec Genie 4
According to the source, the demonstrator with the Apollo robot directly points to future uses in robotics. Being able to maintain a coherent world while injecting changes on the fly allows testing richer scenarios than with simple generative rendering. In this context, Genie 4 looks less like a gadget and more like a research building block on perception, action, and storytelling in simulated spaces.
Google is not the only one pushing this type of tool, but it advances with a particular asset: its ability to link its world models to its Gemini infrastructure. It remains to be seen how these worlds will behave outside demos. AI worlds now speak, but they will also have to hold up when put to the test — which, in research, always happens faster than expected.
Key points
- Genie 4 generates its own sound, according to Google DeepMind.
- Characters can hold live conversations.
- The model is in research preview for trusted testers.
- Access planned next for AI Ultra and Gemini API.
- A demo shows an Apptronik Apollo robot playing checkers.
- Announcement published on October 28, 2026, by The Rundown.
By the numbers
- October 28, 2026 — date of the announcement relayed by The Rundown.
- 2 access categories — AI Ultra and Gemini API partners after the preview.
- 1 robotic demonstration — an Apptronik Apollo playing checkers.