Project Genie: how DeepMind shipped Genie 3 to consumers
Project Genie is the productised Genie 3. You type a prompt, Genie builds a 3D environment you can navigate in real time, and your movement generates the next frame auto-regressively. Early access sits inside Google AI Ultra in the United States (18+), with international rollout staged through 2026.
The current constraints are instructive. Sessions are capped at 60 seconds. Spatial memory lasts roughly one minute — the model remembers objects and state when you turn your back, but not indefinitely. Multiple agents struggle to coexist in the same world. Text rendering is still unreliable. Real-world locations are not reliably accurate. Genie will not yet replace Unreal Engine.
But the frame rate, the persistence, and the fidelity of physics inside that 60-second window are what changed. Genie 3 is the first general-purpose interactive world model to run at 24 fps in real time with coherent physics — a bar that Genie 2, Sora 1 and every prior text-to-video model could not clear.
What Genie 3 actually does, in 2026 terms
A world model is not a video model. A video model produces a clip — a fixed sequence of frames you watch. A world model produces an environment — a state machine that responds to inputs. Genie 3 generates each next frame conditioned on your most recent action, the world's history, and the original prompt, making the output closer to a playable engine than a rendered video.
In practical terms that means three things. First, branching: the same prompt produces a different experience for every user depending on what they do. Second, interaction: you can open a door, walk around an object, and (within the memory window) come back to find it where you left it. Third, transfer: an agent trained inside a Genie 3 world can export learned policies to a physical robot — which is why DeepMind frames the system as an AGI stepping stone, not an entertainment product.
The $10B world-models race: Sora 2, Runway GWM-1, World Labs, Odyssey
Genie 3 is not competing in a vacuum. The 2026 landscape:
- OpenAI Sora 2 — extended consistency, native audio, and interactive editing features that push Sora from clip generator toward interactive canvas.
- Google Veo 3.1 — DeepMind's cinematic-realism sibling to Genie, with synchronised audio and tighter creative controls, aimed at production studios.
- Runway GWM-1 — three-branch world model (Worlds, Robotics, Avatars) targeting gaming, robotics simulation and conversational digital humans. Runway raised $315M from General Atlantic and Adobe Ventures on the back of the launch.
- World Labs (Marble) — Fei-Fei Li's lab shipped its first commercial product with $1B in new funding from Emerson Collective, AMD and Nvidia.
- Luma AI — raised $900M specifically to build world models for advertising, gaming and entertainment.
- Decart — pivoted from its Minecraft-style Oasis demo toward world models for filmmaking, after $100M+ raised.
- Odyssey — streams interactive video frames every 40–50 milliseconds, positioning itself as the 'interactive film' company and courting advertising and travel use cases.
The capital total across the category cleared $10B in Q1 2026. What matters for your planning is the speed of specialisation: within nine months of Genie 3's research release, the field has already fragmented into robotics, avatars, advertising, film and autonomous driving verticals.
From research demo to enterprise rail: Waymo, robotics and simulation
Waymo's February 2026 adoption of Genie 3 is the most important real-world signal. Autonomous-driving simulation used to require bespoke physics engines and months of scenario authoring. Waymo's Waymo World Model — a specialised Genie 3 derivative — can generate physically-consistent traffic scenes on demand and let an agent drive through them, collapsing that pipeline.
Robotics is moving the same way. Training a robot arm in a Genie 3 simulation and transferring the policy to hardware is cheaper, faster and safer than real-world rollouts. The same logic applies to retail floor simulation, warehouse layout testing, and any physical environment where experimentation is expensive.
For enterprises outside robotics, the immediate opportunity is training, onboarding and product demos — any scenario where 'watch a video' could become 'walk through an environment'.
What real-time world models mean for marketers
If you produce content for a living, the mental model shift is this: campaigns stop being linear artefacts and start becoming parameterised experiences.
A product launch video becomes a walk-through of the product's environment. An e-commerce landing page becomes a navigable demo room. A training module becomes a simulated workplace. A brand film becomes a branched narrative the viewer shapes. None of this is available through Project Genie today — 60-second sessions and 720p are not shippable production assets — but the trajectory is legible, and the teams that get fluent with prompt-to-world authoring in 2026 will dominate 2027.
The second-order effect is measurement. World models produce per-user interaction traces by definition. Every session is a clickstream through a 3D space. That data is richer than any video analytics suite — and it is exactly the kind of input a modern marketing stack will be expected to ingest, route, and act on.
Which is the part most teams will not be ready for. Running one Runway render, one Sora clip or one Project Genie session is not hard. Orchestrating prompt-to-world pipelines alongside your brand kit, your ad pixels, your CRM and your analytics — at production cadence — is where it breaks. That is where Anjin comes in.
Anjin: the Marketing Operating System for a world-model era
Anjin is the Marketing Operating System built for a world where the unit of content is no longer a clip but a scene, a simulation or a branched narrative. It sits above the models — Genie 3, Sora 2, Runway GWM-1, Veo 3.1, whatever ships next — and runs the orchestration layer your team needs to use them at scale: brand-consistent prompting, asset versioning, publish workflows, analytics, and the glue between AI output and the channels you already sell through.
Most 'AI for marketing' tools are point features — a caption writer, a one-off video generator, a landing-page autopilot. Anjin is the operating layer underneath those tools, designed so that when the generation frontier jumps again in Q3 2026 (and it will), your team does not need to rebuild their workflow around the new model. You swap the model; the OS stays the same.
Agencies are our launch audience because they feel the pain first. In-house marketing teams feel it next. The category is moving faster than staffing can keep up with — the £888 lifetime license is how you close the gap without hiring a prompt-engineering function you do not have.
The £888 Lifetime License — Offer Closing Soon
Lifetime access to Anjin for a one-time payment of £888. Not a subscription. Not a seat. Not a trial. One payment, unlimited use, for as long as Anjin exists.
The average marketing team spends £888 in about three working days on tooling, freelancers and coordination software. You're buying the platform that replaces most of it — once.
This price will not be offered again once we close our early-access cohort.
Claim your £888 Anjin lifetime license →Founders, agency owners and in-house marketers — this is how you run marketing at AI speed without the team, the burn, or another year of waiting.
Sources: DeepMind, Google Blog, TechCrunch, The Register, PYMNTS, PitchBook, MLQ.ai, Business Standard
