DeepMind unveils Genie 4 world model promising more realistic physics and humans

Google DeepMind has unveiled Genie 4, the successor to the Genie 3 world model that last year rattled gaming stocks when it was revealed. The company says the new model improves on its predecessor in nearly every respect. A limited group of trusted testers has been working with it, and access will begin expanding soon, though no specific timeline has been given.
A world model generates a virtual environment frame by frame in real time rather than requiring pre-rendered scenes or hand-programmed interactive elements. The user describes what they want, and the model builds the experience as it is explored. DeepMind frames this as a step toward artificial general intelligence, and says the technology could supply effectively unlimited training environments for physical robots or speed up prototyping in creative fields. That usefulness depends on how faithfully the model simulates what is asked of it.
Genie 4's improvements draw partly on Google's newer Gemini models. When a world is prompted into existence, Gemini 3.8 Flash acts as a guiding agent that breaks down goals and evaluates each output frame, which keeps the simulation on track so a user can travel through changing environments without losing the original context. A limitation remains: DeepMind notes that Genie 4 worlds stay consistent for only a couple of minutes before they lose track of how things started.
As with Genie 3, parameters can be adjusted mid-simulation. Weather can change, elements can be added and the camera can be moved, with Gemini helping steer the simulation after each update. This resembles how reasoning models generate intermediate internal prompts to converge on the right output tokens.
DeepMind says the physics have reached a new level of realism. Genie 4 handles aerodynamics and collisions better, which the company says makes simulations both more useful and more visually convincing. People depicted in the simulations are said to move and respond to the changing environment more naturally, reducing the uncanny quality common in AI-generated video.
Genie 4 is also the first DeepMind world model with native audio. Generated videos include ambient background sound, and audio elements can be specified in the prompt. The more ambitious claim is human speech: DeepMind says the model can produce interactive dialog that adapts as the simulation changes. In standard AI video, convincing speech marked a major leap in realism, and DeepMind is positioning a similar shift for world models.
Availability is limited and gated. Genie 3 never went beyond the restricted Project Genie web experience, and DeepMind is careful to describe world models as an active research area. Google plans to offer Genie 4 to developers through the Gemini API and AI Studio, but pricing has not been announced. Consumers will get access in phases tied to Google's most expensive AI Ultra plans, which start at $100 per month, with possible expansion to AI Pro and AI Plus plans later. Consumer access will run through Project Genie, which will cap what users can do, and there is a waitlist for early access.
A few implications stand out. The prompt-and-steer model means simulated worlds are produced on demand rather than authored and stored, which changes the cost structure from storage and asset pipelines to inference-time compute. The Gemini 3.8 Flash agent loop suggests each frame carries per-frame model overhead, so latency and price will likely be set by that loop rather than by raw generation alone. A likely trade-off is that tighter steering improves coherence but adds cost and delay, and the roughly two-minute consistency window means anything requiring persistent state must be designed around resets.
Why it matters: Developers building robotics training environments, synthetic data pipelines or interactive prototypes may soon get a more physically plausible simulator with audio and dialog, but they will face an unfamiliar cost model and a short consistency horizon. Gaming and creative-tool teams should watch both the API pricing and the pace of the phased rollout, since the consumer tier is effectively reserved for the $100-per-month AI Ultra plans for now. Whether native interactive speech holds up under real use is the claim most worth verifying independently, since it is the model's biggest announced step and the least demonstrated.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Genie 4 will roll out to developers and Google AI Ultra subscribers soon.