Google DeepMind announced Genie 4 on October 27, 2026, calling it a new generation of real-time interactive world simulation. The model generates photorealistic scenes with native audio (ambient sound, music, and interactive dialogue) and lets users steer weather, camera, people, and objects with natural-language prompts while Gemini models help keep the scene coherent.
Key Facts
| Details | |
|---|---|
| Announcement | October 27, 2026 (DeepMind blog) |
| What it is | Real-time interactive world model; DeepMind's first with native audio |
| Steering | Natural-language edits mid-simulation; Gemini adapts prompts and monitors scene continuity |
| Gemini role | Gemini Omni for real-world appearance/behavior; Gemini 3.8 Flash cited as in-loop guiding agent; experiments also use Gemini 4 Argon with Genie for robotics-style task breakdown |
| Access now | Research preview for trusted testers |
| Next | Coming soon to Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio; waitlist for Project Genie |
| Safety | Safety evaluations before preview; SynthID watermarking on generated video and audio |
| Limitations (DeepMind) | Detail/object drift over longer runs; gaps in complex physics, real-time audio accuracy, legible text, spatial accuracy, long-term consistency |
Background
Genie 3 launched as a research preview a little over a year earlier and stayed limited. Genie 4 adds free-form mid-run edits, stronger physics and human motion claims, and native audio. TechCrunch reports DeepMind said Genie 4 can hold a consistent scene for up to about two minutes, roughly double Genie 3; DeepMind's own blog stresses short-session coherence and lists drift as an open research problem. Ars Technica similarly notes worlds remain consistent for a couple of minutes before context slips.
DeepMind frames hybrid world models (Gemini for structure/planning, Genie for pixel-level look, sound, and response) as the near-term research path, including early Gemini Robotics experiments such as simulating Apptronik's Apollo humanoid playing checkers.
Trusted-tester cohorts span creative direction, archival photo immersion, game prototyping, and enterprise scenario generation. DeepMind says commercial product shape is still early.
Reactions and Impact
TechCrunch quotes DeepMind engineer Diego Rivas on real-time flow for ideation and storyboarding, and research director Tejas Kulkarni on hybrid models that combine Gemini code/text with Genie simulation. Ars Technica notes consumer access is expected first through Google's AI Ultra tier (Ars cites Ultra starting at $100 per month) via Project Genie, with developer access planned through the Gemini API and AI Studio; DeepMind has not published Genie 4 API pricing in the announcement post.
What We Don't Know Yet
- Exact calendar dates for Ultra / API / AI Studio availability beyond "coming soon" / "in the coming weeks" (TechCrunch)
- Genie 4 API or Ultra usage pricing (Ars: pricing not yet available)
- Whether access will later expand to AI Pro or AI Plus (Ars speculation; not in the DeepMind post)
- Independent third-party benchmarks of physics, dialogue quality, or session length outside DeepMind demos