Models

Genie 4 reaches Project Genie, bringing live world-editing to AI Ultra users

Google opened Genie 4 to eligible AI Ultra subscribers on November 18, 2026, adding mid-session world editing, native audio and a community gallery — while flagging scene drift as the key limit.

Google Brings Genie 4 to Project Genie for Eligible AI Ultra Users
Google Brings Genie 4 to Project Genie for Eligible AI Ultra UsersVitaly Gariev / unsplash
By James Calloway5 min read

Why it matters

  • Google opened Genie 4 to eligible AI Ultra subscribers on November 18, 2026, following the October 27, 2026 research preview.
  • Genie 3's per-generation limit rose from 60 seconds to three minutes.
  • Gemini 3.8 Flash acts as an in-loop guiding agent that observes frames and replans.
  • All outputs carry SynthID watermarking and pass safety screenings.
  • Google cites scene drift, camera latency and object permanence as unresolved limits.

Google opened Genie 4 to eligible AI Ultra subscribers on November 18, 2026, giving them first-time access to the world-simulation model inside the browser-based Project Genie prototype. The rollout pairs the model with a conversational creation agent and a community gallery for publishing and remixing worlds. Google AI Pro members can join a waitlist but are excluded from the initial wave.

The announcement executes part of a plan Google DeepMind laid out on October 27, 2026, when it put Genie 4 into research preview for trusted testers. At that time, DeepMind named Ultra subscribers in Project Genie and early-access partners on the Gemini API and Google AI Studio as the next release groups. The Project Genie piece has now arrived.

What can Genie 4 do inside a session?

The headline capability is mid-session control. Users can type prompts while navigating a world to change the weather, add a character or object, move the camera or redirect events — without restarting the simulation. The original Project Genie prototype, powered by Genie 3 and rolling out since January 29, 2026 to U.S. AI Ultra subscribers aged 18 and over, supported sketching worlds from text and images, navigating them and remixing creations. Editing a running world mid-flight is the new step.

Genie 4 also generates native audio alongside visuals: ambient sound, music and character dialogue that react to user inputs and events in the simulation. Google positions it as the higher-fidelity option for visuals, physics and audio. Genie 3 remains selectable, and it now has a clear niche — longer runs and worlds seeded from 360-degree panorama images. Its per-generation limit has been raised from 60 seconds to three minutes, and users can chain multiple worlds into longer sequences or trigger real-time events with natural-language prompts.

The takeaway: Google is treating world simulation as a controllable medium, not just a video generator — but the older model still carries the workloads the new one cannot hold together.

How do Gemini models keep the simulation coherent?

Google DeepMind describes Genie 4 as natively multimodal, with Gemini Omni (the omni-modal tier of Google's Gemini model family) supplying real-world reference for how simulations should look and behave. Other Gemini models contribute reasoning for coherence and continuity.

One of them, Gemini 3.8 Flash, can operate as an in-loop guiding agent. It observes the scene, decomposes goals into steps, checks incoming frames for progress and revises its plan when the world or the user's input shifts. That is a closed feedback loop: the guide watches what the simulation actually renders and corrects course, rather than planning blind.

DeepMind frames this coupling as a way to model multi-step events and to test how AI systems navigate environments before real-world deployment. Its stated use cases are training environments for robots and agents, plus new interactive experiences. That ambition deserves scrutiny. A simulator whose object permanence drifts over long sessions is a shaky proving ground for anything safety-relevant, and DeepMind itself has flagged that gap — more on that below.

The takeaway: the Gemini-in-the-loop architecture is the substantive research claim here, and it is exactly the part Google cannot yet guarantee will hold up over time.

What does the creation agent and gallery add?

The Project Genie agent handles initial creation through natural-language conversation. It coordinates Google's multimodal models to produce a world's opening visuals and a custom background score matched to the intended mood. During exploration, Gemini can watch the screen, suggest context-aware next actions through on-screen chips and trigger events automatically.

The gallery adds a publishing and reuse layer. Users can modify an existing world's structure, soundtrack or rules with the agent, then publish their version for others to explore and remix. Google cites two examples:

  • Living Memory, which converts personal photographs into explorable spaces.
  • The Life of Ants, created with filmmaker Anand Gandhi, which presents a colony from an ant's perspective.

These are curated showcases, not evidence of what average sessions produce. Google has not published figures on generation quality, latency budgets or how the remix layer handles attribution at scale.

The takeaway: the social layer is where Google hopes worlds become a creative medium, but the demonstrated range so far rests on hand-picked examples.

What are the safety measures and the honest limits?

Every video and audio output in Project Genie carries SynthID, Google's imperceptible watermark for identifying AI-generated content, and all model inputs and outputs pass safety screenings, the company says.

Google is unusually specific about what does not work. It calls both Genie models frontier research systems and lists camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence as unresolved challenges. In Genie 4 specifically, visual details and object permanence can drift during longer sessions, after complex edits or while navigating. Legible text and spatial accuracy remain active research areas.

That candor matters because it defines the boundary of the launch. The expansion increases who can build with Genie 4, but the durability of its simulated worlds is the binding constraint — both on long creative sessions and on DeepMind's proposed robot and agent training environments. A world that forgets its own objects mid-run cannot yet serve as a reliable stand-in for physical reality.

The takeaway: the watermarking and screening are table stakes; the drift problem is the open engineering question that will decide whether Genie becomes infrastructure or stays a demo.

Why this rollout matters

The November 18 release converts an October research preview into a product surface for Google's highest-paying AI tier, and it maps Google's competitive position in generative interactive media: OpenAI and others have shipped video generation, but few have shipped persistent, editable, audio-native worlds with an reasoning model supervising the loop.

The staged access also signals sequencing. Ultra subscribers first, API and AI Studio partners next, Pro members on a waitlist — Google is widening the funnel slowly, consistent with a system it still labels research-grade. Until scene permanence and latency improve, expect Genie 3's three-minute generations, not Genie 4's fidelity, to anchor the longest projects.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Google DeepMind's Genie 4 simulates worlds in real time — with sound and talking characters
  2. Google DeepMind ships Genie 4: photorealistic worlds, audio, and real-time steering
  3. Google Opens Project Genie to AI Ultra Subscribers in the U.S.
  4. Google Ships Gemini 4 Carbon and Genie 4, Its Second Frontier Release in Two Weeks
  5. Google Makes Tasks in Gemini Free Across Six Countries

« Previous articleNext article »