Separating AI hype from reality.

Google DeepMind Rolls Out Genie 4 to Ultra Subscribers, Calls It a Research Preview

November 18, 2026 · Google · Research

Google DeepMind has begun rolling out Genie 4, its latest generative world model, to eligible Google AI Ultra subscribers through the Project Genie browser experience. The company describes the release as a research preview for trusted testers, with broader consumer and developer access to follow [3].

What Shipped Today

Genie 4 is accessible only to Google AI Ultra subscribers via the Project Genie website. Google AI Pro members can join a waitlist. The release includes:

  • Genie 4 model: real-time world generation with mid-session editing via text prompts and native audio (ambient soundscapes and character dialogue) [1][2].
  • Updated Genie 3: session length extended to up to three minutes (from 60 seconds), world chaining, 360-degree panorama grounding, and real-time in-world events [1][2].
  • Project Genie agent: a conversational interface that orchestrates Google's multimodal models to build initial visuals and soundtracks, then suggests or triggers events during exploration [1][2].
  • Community gallery: curated and user-created worlds that can be explored, remixed, and published. Launch collaborations include "Villa Giallo" with creative technologist Will (BLVCKL!GHT) and "The Life of Ants" with filmmaker Anand Gandhi [1][2].

All generated video and audio carry SynthID watermarks, and inputs and outputs pass through safety screenings [1][2].

What the Company Claims

DeepMind's research blog announces Genie 4 as a new generation of world simulation and says it is natively multimodal, generating audio and visuals simultaneously. It says Genie 4 incorporates Gemini models: Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models provide reasoning for coherence and continuity [3].

The blog highlights four capability areas:

  • Real-time shaping: natural-language prompts mid-session to change weather, add characters, or move the camera, with Gemini adapting prompts to scene context [3].
  • Physics and human action: simulations of aerodynamics, collisions, and characters that "move naturally, respond organically to their environments, and can be interactively steered in real-time" [3].
  • Native audio: ambient sound and interactive dialogue that adapts on the fly [3].
  • Complex event simulation: Gemini 3.8 Flash acts as a guiding agent, observing frames, breaking goals into steps, verifying progress, and adapting plans as the world shifts [3].

The post includes numerous video examples — saxophone player on a floating platform, pink fabric river in a meadow, hot-air balloon over Antarctica, watermelon on a plinth, maple seed falling, 1980s Japanese train station, stained-glass world, dragon-racing first-person view, motorcycle scene, robot pipette task, porcelain gallery, and a magician with a peacock — all presented as real-time interactions [3].

What Is Actually Demonstrated

The videos are embedded in the company's own blog posts. The provided announcements do not include independent benchmarks, third-party evaluations, or public API documentation. The demo prompts are highly detailed, often hundreds of words long, specifying camera behavior, lighting, film stock, and negative constraints (e.g., "Absolutely no HDR, no dramatic contrast, no glossy digital rendering") [3]. The announcements do not show performance with shorter, less prescriptive prompts.

Self-Acknowledged Limitations

DeepMind explicitly states that "Genie 3 and Genie 4 remain world models in the frontier research space" and lists active challenges: "camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence" [1][2]. A separate Genie 4 model research overview is referenced for deeper limitations [1][2].

Access and Unknowns

  • Pricing: Google AI Ultra subscription cost is not stated in the announcements.
  • API access: The research blog mentions "early access partners of the Gemini API and Google AI Studio" but gives no timeline for general developer access [3].
  • Model weights: Not released; the models run in-browser via Project Genie.
  • Performance metrics: No frame rates, latency numbers, or compute requirements are disclosed.
  • Comparison: No benchmarks against competing world models (e.g., from NVIDIA, Runway, or open-source projects) are provided.

Context

All three sources are Google/DeepMind owned channels. There is no independent reporting, academic paper, or third-party testing in the provided material. The rollout is gated behind a premium subscription and a waitlist, limiting the pool of users who can verify the claimed capabilities today.

What happens next — broader availability, API launch, independent evaluation, or further research iterations — remains to be seen.

Sources

  1. Genie 4 is now available (blog.google)
  2. Project Genie update featuring Genie 4 (deepmind.google)
  3. Genie 4: A new generation of world simulation — Google DeepMind (deepmind.google)