AI DevelopmentNew Release8 min readPublished

Genie 4 in Project Genie: Who Gets Access and What Works

Genie 4 lets you change a world while you move through it and hear the scene respond. The older model gives you more time in that world.

DA
Digital Applied
PublishedNovember 19, 2026
Read time8 min

If you want to hear a market square fill with rain as you walk through it, choose Genie 4. If you want a viewer to spend three minutes exploring the square from a panoramic photograph, choose Genie 3. Google’s November 18 update puts both models in Project Genie, and the older one still does several things the new release does not promise. Google’s Project Genie announcement

The other easy mistake is buying the wrong Ultra plan. Google’s announcement describes a rollout to eligible subscribers. The Google One plan page puts the actual restriction in a footnote: Project Genie comes with Ultra 20x. Someone on Google AI Pro can join a waitlist. For a creative lead deciding whether to pay for a month, the useful question is what a short, interactive session will settle that an ordinary storyboard cannot.

Key takeaways
  1. 01
    Check the exact plan.Google One names Ultra 20x as the qualifying tier. Google Labs requires a personal Google Account, age 18 or older and a supported location. Pro members can request access.
  2. 02
    Choose Genie 4 for a live scene change.Direct a change in weather, action or character while navigating and hear native sound respond. DeepMind describes around a minute of continuous interaction.
  3. 03
    Choose Genie 3 for a longer route.Its November upgrade allows up to three minutes per generation, linked worlds and 360-degree panorama input. It can also respond to in-world events.
  4. 04
    Save the playthrough you intend to show.A replay can produce a different outcome. Google lets you download a video of the session, share the experience or publish a version to its gallery.

01 — AccessCheck the Ultra tier

Google Labs says Project Genie requires a personal Google Account and a user aged 18 or older in an eligible country. Its help page describes entry for Ultra subscribers and a gradual waitlist for Pro users. Google One specifies which Ultra tier counts in the United States: its Project Genie footnote says Ultra 20x, with availability in more than 140 countries. The cheaper Ultra 5x tier shares the Ultra name. Its name alone is no reason to upgrade for Genie. Google One’s plan comparison and Project Genie footnote

For a US buyer comparing monthly subscriptions, Pro is $19.99 and Ultra 5x is $99.99. The qualifying Ultra 20x tier costs $199.99 per month, exactly $180 more than Pro at standard US pricing. Someone who already uses 20x for other work faces a different decision: they can try the prototype as part of the subscription. Check the country and plan shown on your own account before budgeting, since local pricing and availability can differ. Google’s US subscription prices

The waitlist gives a Pro member somewhere to register interest without spending that difference. Google says access is granted gradually and confirmed by email. Daily use varies with plan and demand, and there can be a queue to create a world. During early access, world generation does not consume the subscriber’s AI credits. Neither that credit policy nor a paid subscription guarantees a slot at a particular time. A person awaiting access can still view short gallery previews. Google Labs eligibility and usage rules

02 — InteractionDirect the next moment

In Project Genie, a world is an experience you can move through and influence. For Genie 4, start with a description or reference image and navigate the resulting scene. While moving, type an instruction that changes something in the world. Google demonstrates changes to weather, characters and action. Genie 4 also generates native audio as the action unfolds, including environmental sounds and character dialogue. This makes sound part of exploring the scene instead of a track supplied after recording. DeepMind’s Genie 4 capabilities

Imagine a sound designer and an art director discussing how to reveal an approaching train. A still can establish the empty platform. A video can show the director’s preferred reveal. An interactive world gives the pair another question to ask: what should someone at the other end of the platform hear or see first? They could navigate away from the track, prompt the arrival and compare the impression with a playthrough facing the train. This is a proposed creative exercise, and the timing of the sound is itself something to evaluate.

DeepMind’s model page describes around a minute of continuous interaction. Google Labs says individual session duration depends on the model and experience. Plan one short sequence with a beginning, one change and an observable consequence. Google’s release gives no fixed Genie 4 output resolution to promise to a client. If a particular delivery format is required, check a downloaded playthrough before making it part of the brief.

03 — Model pickerWhen the older model fits better

Genie 3 now supports up to three minutes per generation, up from 60 seconds. It can connect worlds into a longer journey and use a 360-degree image to ground a space. Google illustrates the panorama feature with Petra, where the canyon and façade provide recognisable landmarks as the viewer turns around. These are named Genie 3 features in the November announcement. Moving the camera through a Genie 4 scene does not give it the published three-minute limit. Google’s model-by-model feature list

For a designer presenting a visitor route, the panorama can be more useful than new dialogue. Start with the panorama at an entrance, trace a path to a landmark and return to the starting view. The reference gives the designer a basis for judging the layout. A second scene can extend the journey. World chaining joins experiences, so a continuous-looking transition still deserves a check if an object or character must carry through from an earlier space.

Both models can respond to prompts during exploration. Genie 3’s update includes real-time promptable events, while Genie 4’s distinguishing combination is live natural-language editing with audio generated as the scene develops. Choose between them by the thing a reviewer must evaluate: sound and immediate reaction, or a longer visit anchored to a supplied view.

Hearing a scene respond matters
Stage one meaningful change in a short scene. Listen for whether the sound follows the action.
Genie 4
The route or location matters
Start from a panorama, allow time to turn and return, and use chaining where the visitor moves between worlds.
Genie 3
The audience only needs to watch
Download the accepted playthrough so each reviewer receives the same sequence.
Saved video

04 — CreationA model and an experience builder

The Project Genie agent sits above the world model. Google says users can describe an experience in conversation and the agent coordinates multimodal models to create its visuals and background score. Gemini can also observe the session, suggest a next action or trigger an event in response to what is on screen. A background score assembled for the experience and native audio arising inside a Genie 4 scene are different parts of the product. A polished gallery example may involve both. Google’s description of the Project Genie agent

The gallery is useful for deciding what to prototype before designing a world yourself. Google’s Dino Field Guide uses panoramic settings. Villa Giallo is a room-by-room mystery, while The Life of Ants places the viewer at ant scale. Pick an example that shares a problem with your own project. For a clue-driven story, whether a viewer can find and recognise the clue on returning to a room matters more than whether the opening shot looks impressive.

Google also describes a Rock Paper Scissors experience where Gemini referees the action. That is a useful distinction for someone estimating engineering work. A model can render a visible outcome, and a separate component can interpret what happened. If a game depends on unambiguous scoring, keep the authoritative state and rules in software that can be checked. A convincing animation alone cannot settle a disputed result.

05 — LimitsFour things to watch

Google calls out camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence as active research problems. They show up in different parts of a critique. A camera that responds late can make a composition feel heavy even when the image is sharp. A character picking up a prop tests contact and motion. A line spoken while a door closes tests whether the voice and impact arrive when the viewer expects. Returning to a room tests whether the prop is still there. DeepMind’s research overview and limitations

Give each of those observations its own note. For the train-platform example, write down the prompt and camera position, then save the moment where the sound is too early or the bench changes shape. A second person can decide whether the problem belongs to the concept or the prototype. A beautiful pass that conceals a timing defect is poor evidence for a cue-driven game, while a slight sound mismatch may be acceptable in an early conversation about lighting.

DeepMind also describes gaps in precise spatial accuracy and legible text. For a concept based on a recognisable place, compare the generated view with the reference photograph. For training material that depends on the reader following a sign or a switch label, a human needs to check the label in the actual recording. Use a controlled simulation or a conventional engine when correctness depends on measured dimensions, stable state or repeatable physics.

06 — HandoffSave the playthrough

A reviewer can receive two quite different things. A downloaded video records the playthrough you approved. A shared experience lets another person explore, and Google says replaying an experience can produce different results. For a creative sign-off, save the video and label the moment the client is being asked to approve. For a discussion about exploration, share the world and specify the action the reviewer should try. Google Labs sharing and download guidance

Sharing also determines who can see the ingredients. When remixing is allowed, other people can see and reuse the prompts, images and assets used to create the experience. Publishing to the gallery exposes the submitted version and the name associated with the creator’s Gmail address. Later edits need publishing again to appear in that gallery version. Use a reference image cleared for the intended audience, especially when a concept relies on an unreleased set design or another person’s likeness.

Google says all generated video and audio carry its SynthID watermark, and both inputs and outputs go through safety screening. That gives provenance information for generated media. It does not grant permission to share a source image or guarantee that a recorded experience reproduces a real location accurately. Keep the rights record and the approved recording with the materials delivered to a client. Google’s safety and watermarking statement

07 — Buying decisionWhat can one session settle?

The strongest reason to buy a month of Ultra 20x is a creative decision that requires someone to act inside a scene. Write down that question first. If it is which way a visitor turns when the light drops, invite the viewer to navigate. If it is whether a blue jacket reads against the background, a still might settle it more cheaply. Put the intended reviewer in front of the prototype while there is still time to use what they learn.

A software team wanting a callable simulation service has another requirement. DeepMind’s earlier research announcement described early-access partners for the Gemini API and AI Studio. The November release delivers the Project Genie web experience. It does not announce a public API, a downloadable world package or a production service guarantee. OpenRouter’s public model catalogue had no Genie listing on November 19. The Genie 4 research preview covers the earlier partner stage.

The Google AI plan comparison covers the other reasons a subscriber might pay for Ultra. The Runway world-model guide gives a separate reference point for teams comparing creative workflows. A brief for an AI prototype engagement can specify the same deliverable: a recorded decision, a playable exploration, or an application requiring a supported API. Those deliverables have different costs.

Decision

Buy the session for a decision you can name

Use Genie 4 to explore a short scene that should respond to movement, sound or a new instruction. Use Genie 3 for the longer or panorama-led route. If the Project Genie waitlist or public gallery can answer your question, the top Ultra subscription can wait.

Digital Applied

Define a test for an interactive prototype

Specify what the viewer should try, what the team must observe and which recorded result the reviewer needs.

Creative briefPrototype scopeReview format
Evaluation

Set the acceptance criteria

  • →One interaction
  • →One observation
  • →One reviewable output
Answers

Frequently asked questions

Google says the rollout is for eligible subscribers. Its plan comparison specifies Ultra 20x for Project Genie, with availability tied to country and account. Google Labs requires an adult user signed into a personal Google Account.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Genie 4: What Interactive Worlds Can Actually Do Now

Genie 4 adds native audio and live control to generated worlds. Understand preview access, creative uses, simulation limits and what to test before buying.

October 27, 2026 · 9 minRead
AI Development

Gemini 4 Carbon: Who Can Use Google's New Cyber Model

Google announced restricted Carbon access in October. Read the original cyber-programme analysis and the November 10 update on public access and pricing.

October 29, 2026 · 11 minRead
AI Development

Gemini 4 Argon: Price, Benchmarks and When You Can Use It

Gemini 4 Argon launches at $2/$10 per million tokens, rising to $4/$20 later. Who can use it now, what Google's benchmarks show and how to prepare.

September 30, 2026 · 6 minRead
AI Development

Gemini Omni 1.1 Flash: 40-Second Scenes, Drafted at 360p

Google's Omni 1.1 Flash chains 10-second extensions up to 40 seconds and documents 1080p and 4K as upscales. What the reported $0.03/s draft tier buys.

August 27, 2026 · 14 minRead
AI Development

Gemini 3.5 Flash Cyber and the Rise of Gated AI Models

Google's Gemini 3.5 Flash Cyber finds and patches vulnerabilities but stays gated. Why restricted access is becoming the default for dual-use AI models.

July 23, 2026 · 11 minRead
AI Development

Agent Computer Use: Enterprise Automation Playbook

Enterprise playbook for deploying computer-use agents — a 40-point guardrails checklist spanning identity, audit, action boundaries, failures, and compliance.

May 22, 2026 · 17 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source