Genie 4: the world-simulation model, read from the source

"Genie 4" is the query for Google DeepMind's newest world model — and it is one of the rare model words where the primary source is unusually rich. The official post, "Genie 4: A new generation of world simulation," went up on DeepMind's blog on October 27, 2026 observed, one day before Google's Gemini 4 Carbon announcement and two days after OpenAI's GPT-6.1 Terra observed. DeepMind's own framing: Genie 4 generates real-time, interactive simulated worlds — with simulated physics, native audio, and people who move, react and speak — and the company states plainly that "understanding and simulating the world is a crucial aspect of building AGI" vendor claim. This page unpacks what the announcement actually says, how access is gated, and what it leaves unsaid. Labels: vendor claim / observed / analysis.

What Genie 4 is, per the announcement

• A world model, not a chat model. Genie 4 belongs to DeepMind's Genie line — the "generate and explore interactive worlds" family on DeepMind's own site navigation observed. You do not prompt it for text; you set an event in motion and it simulates what happens next: how the world responds, how it sounds, and how the people in it move and speak vendor claim.
• Real-time interactivity. The headline advance over prior Genie generations is real-time: you steer while the world is being generated, with keyboard controls for navigation and camera and text prompts to steer the scene vendor claim.
• Simulated physics. DeepMind says Genie 4 is "built on an advanced physical understanding of the world" and simulates scenarios like aerodynamics and collisions — a claim about consistency of the simulation, not a physics-engine datasheet vendor claim.
• Native audio and natural human motion. Sound is generated with the world, and simulated people "move, react, and speak" vendor claim.
• Real-time editing. You can modify a running simulation with natural language — change the weather, move the camera, introduce a new person or object — and the system adapts the prompt contextually vendor claim.

The part most coverage skips: Gemini in the loop

The announcement's most operationally interesting detail is architectural: Genie 4 does not run alone. A Gemini model "serves as a guiding agent in the loop" — continuously observing the scene, breaking down goals into a sequence of steps, and reviewing incoming frames to verify each step succeeded, adapting the plan as the world shifts or the user intervenes vendor claim. DeepMind credits this partnership for "reliable causality, procedural behaviors, and multi-step reasoning" in Genie 4 worlds vendor claim. The pattern to notice: world models are becoming systems — a generator plus an overseer model — the same composition direction the rest of the industry is taking analysis.

Model card (from the official announcement)

Field	What the announcement says
Name	Genie 4 observed
Maker	Google DeepMind observed
Announced	October 27, 2026 observed
Category	Real-time interactive world simulation (world model) vendor claim
Key stated capabilities	Real-time generation; simulated physics (aerodynamics, collisions); native audio; natural human motion and speech; prompt and keyboard steering; real-time natural-language editing vendor claim
Architecture detail	Gemini model as guiding agent in the loop (goal decomposition, frame review, plan adaptation) vendor claim
Access	Gated: researcher/security-expert testing → Google AI Ultra subscribers via Project Genie + early-access Gemini API / AI Studio partners; waitlist vendor claim
Pricing / specs	Not stated in the announcement we captured analysis

Access: who gets in, in what order

DeepMind's stated sequence vendor claim: security testing with researchers and experts comes first; release to consumers and developers follows, "starting with Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio," with a waitlist for Project Genie access observed. Read that as a three-gate funnel — lab testing, subscription tier, API partners — before anything resembling general availability analysis. If you are not an AI Ultra subscriber or an early-access API partner, the honest status is: not yet, and no published date analysis.

What the announcement does not say

• Pricing. No token rate, subscription requirement detail, or API price analysis.
• Technical specs. No resolution, frame rate, session length, or context figures analysis.
• Benchmarks. No quantitative world-consistency or physics evaluations are published with the post analysis.
• General availability date. Waitlist now; no calendar analysis.

Details await official documentation. World-model evaluation is younger than language-model benchmarking — there is no SEC-bench equivalent for "did the simulated teacup fall believably" — so expect claims here to stay qualitative longer analysis.

Why world models matter beyond the demo

• Training ground for agents and robotics. DeepMind's own site ties Genie to robotics simulation ("an infinite playground" in its model listing) vendor claim; simulating environments lets AI systems be tested "before bringing them to the real world," per the announcement vendor claim.
• Safety testing at scale. "Safe, infinitely variable simulations" is DeepMind's stated use — generating many controlled scenarios to probe how systems behave vendor claim.
• The System One angle. A world model that reacts consistently is, in effect, a decision environment you can query fast — adjacent to the fast, structured-decision models this site tracks (see What is a System One Model?) analysis.

Related pages

• Gemini 4 Carbon — Google's frontier model, announced the day after Genie 4
• Gemini 4 Carbon vs GPT-6.1 Terra — the same-week cross-vendor comparison
• What is a System One Model? — the fast structured-decision category this site tracks
• Migos AI video — generative video for creators, a different slice of the media-model wave

Genie 4 FAQ

Is Genie 4 a game engine?

No — DeepMind positions it as a world model for research, robotics simulation and interactive exploration vendor claim. It may look game-like in demos, but there is no stated game-dev toolchain or engine integration in the announcement analysis.

How is Genie 4 different from Genie 3?

The announcement's own emphasis: real-time interaction, simulated physics scenarios like aerodynamics and collisions, native audio, and the Gemini guiding-agent loop vendor claim. We avoid inventing a version-by-version spec table — prior-generation comparisons should wait for DeepMind's own documentation analysis.

When can developers call Genie 4 through the Gemini API?

Early-access partners of the Gemini API and Google AI Studio are in the second wave of DeepMind's rollout plan vendor claim. No public API date is published analysis.

Is there a waitlist?

Yes — the announcement links a waitlist for Genie 4 in Project Genie (Google AI Ultra) observed. Getting on a waitlist is not getting access; treat availability claims accordingly analysis.

Will Genie 4 replace video-generation models like Veo?

Different jobs: Veo generates cinematic video clips; Genie 4 generates persistent, steerable, interactive worlds vendor claim. Overlap at the demo level, divergence at the product level — but convergence is a prediction, not a fact analysis.

This page is descriptive and independent, built from a direct capture of Google DeepMind's official announcement, last reviewed November 1, 2026. Not affiliated with Google or DeepMind. "Genie", "Gemini" and "Veo" are trademarks of Google; vendor claims are labeled and not independently verified.