Google Genie 3 Review: The First Playable 'World Model'

While the world was distracted by video generation wars between OpenAI's Sora and Runway, Google DeepMind just quietly changed the game entirely. Meet Genie 3, an AI that doesn't just make videos, it builds playable, interactive worlds.

Released as a research prototype under "Project Genie," this new model represents a fundamental shift in generative AI. It's not just predicting the next pixel in a frame; it's understanding physics, collision, and user intent to create a consistent 3D simulation you can actually control.

What is a World Model?

At its core, Genie 3 is a "World Model." Unlike a video generator that creates a passive MP4 file, a world model simulates an environment. It understands that if a character jumps, gravity should pull them down. If they walk behind a wall, they are occluded but still exist.

DeepMind's goal isn't just entertainment. By creating diverse, simulated environments, they are building the training grounds for future Artificial General Intelligence (AGI) agents. An AI agent can't learn to navigate the real world safely if it only ever watches videos of it, it needs to experience it. Genie 3 provides that experience.

How Genie 3 Actually Works

Under the bonnet, Genie 3 is an auto-regressive model: it generates the world one frame at a time, conditioning each new frame on the world description and the growing history of your actions. Because you can press a key several times per second, the model has to produce each frame in real time while staying coherent with everything that came before, which is a substantial computational feat at 720p and 24fps.

The most surprising design decision is what Genie 3 does not have: an explicit 3D representation. Techniques like NeRFs and Gaussian splatting build a fixed geometric model of a scene and then render views of it. Genie 3 instead conjures every frame from scratch, and DeepMind reports that its environmental consistency is an emergent property of the model rather than something engineered in. The trade-off is that its worlds are far more dynamic and malleable than a pre-baked 3D scene, at the cost of occasional physics wobbles.

• Visual memory: The model retains up to a minute of prior visual context, so if you paint a wall and walk away, the paint is still there when you return.
• Promptable world events: Beyond movement controls, the research model accepts text prompts mid-session to change the weather or introduce new objects and characters, enabling "what if" counterfactual scenarios.
• Session horizon: Interactions remain coherent for a few minutes of continuous play, a dramatic improvement over the seconds-long horizons of earlier world models.

DeepMind is candid about the current limits: the range of actions an agent can take is still constrained, interactions between multiple independent agents are poorly modelled, real-world geography is simulated inaccurately, and legible in-world text rarely survives generation.

From Video to Playable Worlds

The define feature of Genie 3 is interactivity. You can input a text prompt, an image, or even a rough sketch, and the model generates a 3D environment. But here's the kicker: you can then control the character or camera within that environment in real-time.

Current capabilities include:

• Real-time Generation: It generates frames at up to 24fps (and 720p resolution), allowing for smooth gameplay.
• Long-form Consistency: Unlike previous models that hallucinate wildly after a few seconds, Genie 3 maintains world state for minutes at a time.
• Multi-modal Inputs: Draw a map on a napkin, scan it, and Genie transforms it into a 2D platformer or 3D dungeon.

Genie 3 vs. Sora & Veo

It's important not to confuse Genie with video generation models like Sora or Google's own Veo.

Release Date & Access

Genie 3 was officially released in August 2025 as a research paper, but the playable "Project Genie" prototype became accessible to select users on January 29, 2026.

Hands-On: What Early Users Found

Early hands-on coverage gives a good sense of what the prototype feels like in practice. TechCrunch's Rebecca Bellan used it to build a claymation-style castle made of marshmallows, complete with chocolate-sauce rivers and candy trees. The pipeline is straightforward: you describe an environment and a character in text, Google's Nano Banana Pro image model (part of the same family as the Nano Banana line we reviewed) converts the prompt into a starting image, and Genie 3 turns that image into a world you can roam in first- or third-person view using the arrow keys, W-A-S-D and the mouse. Generation takes a matter of seconds.

The most talked-about restriction is the 60-second session cap, which is a compute decision rather than a model limit. As DeepMind's Shlomi Fruchter told TechCrunch: "The reason we limit it to 60 seconds is because we wanted to bring it to more users."

• Where it shines: Stylised worlds, including watercolour, anime and cartoon aesthetics, generated impressively, and the model showed genuine memory consistency when returning to previously visited spots.
• Where it struggles: Photorealistic and cinematic prompts tended to look like video games rather than life, controls were occasionally unresponsive or sent the character the wrong way, and characters sometimes clipped straight through walls.

Fruchter was refreshingly honest about where this sits: "We don't think about [Project Genie] as an end-to-end product that people can go back to everyday, but we think there is already a glimpse of something that's interesting and unique."

Mid-2026 Update: Street View Worlds & Global Rollout

Project Genie has moved quickly since its January debut. In May 2026, around Google I/O, DeepMind shipped the prototype's first major expansion, and the headline feature is Street View grounding: tap the Maps pin, choose a real place in the US, optionally pick a style such as "Desert Sands" or "Stone Age", and Genie generates an explorable world anchored to Street View's real imagery. The feature is powered by Maps Imagery Grounding, drawing on an archive of roughly 280 billion Street View images spanning 110 countries.

The same update added two quality-of-life features that push Genie from tech demo towards platform:

• A library: Generated worlds can now be stored, organised and revisited, and past creations can be remixed to iterate on ideas.
• External sharing: Worlds can be shared outside the app. Anyone can watch a video preview, while Ultra subscribers can step in and explore or remix them.

Just as significantly, access stopped being US-only: Project Genie began rolling out globally to eligible Google AI Ultra subscribers aged 18 and over, on Google's $200-per-month (approx £160) top tier. Google continues to describe it as an experimental research prototype in Google Labs, with accuracy and visual fidelity still being refined, and tracked AI channels on YouTube have kept the updates in steady circulation through the summer.

Genie 3 and the Road to Embodied Agents

The consumer toy is arguably the sideshow. DeepMind's own research demonstrated Genie 3's real purpose by dropping its SIMA agent into generated worlds and instructing it to pursue goals, such as walking to specific objects and locations in a warehouse setting, which the agent achieved across extended action sequences. The promptable-events capability makes this even more powerful: researchers can inject rare or dangerous scenarios into a world on demand and see how an agent copes, without any real-world risk.

The follow-up was more striking still. SIMA 2, the Gemini-powered generalist agent DeepMind announced in November 2025, has been shown improving at tasks inside Genie-generated worlds without new human-supplied training examples: one AI inventing the worlds in which another AI practises. If that loop keeps compounding, world models stop being a content feature and become the training infrastructure for embodied AI, from game NPCs to household robots. That, far more than marshmallow castles, is why DeepMind is spending so much compute on this.

Verdict

Genie 3 is a glimpse into a future where "content" isn't just something you watch, but something you inhabit. For game developers, this is the holy grail of infinite procedural generation. For AI researchers, it's a vital step toward embodied AGI. For the rest of us? It's just incredibly cool magic.