Google's Genie 4 world model adds sound, speech and longer scenes
Google DeepMind put Genie 4 into research preview on Tuesday, and it is the first real-time world model that generates the soundtrack along with the scene. The number worth remembering is two minutes: that is how long Genie 4 can hold a consistent world, double what Genie 3 managed.
Genie 4 generates audio, interactive dialogue and physical motion at the same time as the pixels, and Google is pitching it as the base layer for simulating anything. The model is natively multimodal, so music, ambient noise and character speech are produced with the world rather than dubbed over it: a lake scene carries the sound of water, and people in the scene answer you in their own voices. DeepMind's announcement groups the gains into four: free-form language steering, better physics, native audio, and multi-step event simulation.
Steering is the part that reads as a genuine change rather than a demo flourish. You can change the weather mid-scene, move the camera, or drop a new person or object into the world, and Gemini adapts the prompt to what is already on screen; between your inputs, it keeps the scene progressing on its own. The physics work covers aerodynamics and collisions, and the biggest visible jump is in people — characters that move, react to their surroundings and can be steered while they are moving.
Under the hood this is a hybrid, not a single model. Gemini Omni anchors the simulation in real-world appearance and behaviour, and other Gemini models supply the reasoning that keeps a scene coherent. In real time, Gemini 3.8 Flash acts as a directing agent in the loop: it watches the generated frames, breaks a goal into steps, checks whether each step succeeded and re-plans when the world shifts or the user intervenes. "What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction," Google research director Tejas Kulkarni said during a briefing, TechCrunch reports.
The concrete numbers are thin on purpose. DeepMind published no benchmark scores, no latency figure and no price. What it did confirm is the two-minute coherence window, the fact that Genie 3 shipped as a preview a little over a year ago, and that every generated frame and audio track carries a SynthID watermark. Availability starts narrow: trusted testers now, then Google AI Ultra subscribers inside Project Genie plus early access partners on the Gemini API and Google AI Studio, "coming soon" with a waitlist in the meantime.
The robotics experiment is the one to watch. DeepMind paired Gemini 4 Argon with Genie 4 so that Argon decomposes a task into steps and feeds them to the simulation as captions while it renders — in the published example, Apptronik's Apollo humanoid working through a game of checkers. That is a plausible route to testing robot behaviour in simulated worlds before committing hardware and floor time, and it is the use case that would justify the research spend if the drift problem gets solved.
Drift is still the problem. DeepMind is unusually direct about it: visual details and object permanence can shift over longer sessions, after several complex edits, or while navigating a scene, and accurate audio, legible in-world text and spatial precision remain open research. Google DeepMind engineer Diego Rivas told reporters it is still early days and that the end-to-end product experience is not settled. Anyone who has watched a generated world melt after ninety seconds will recognise the caveat as the real ceiling, not a footnote.
The commercial context explains the timing. World models are consolidating fast: AMD agreed to buy Fei-Fei Li's World Labs for $8.2 billion in September, Amazon has reportedly been in talks to buy Decart for around $7 billion, and Rivian bought Third Dimension this month to bring driving simulation in-house. Runway shipped its second playable world model, GWM Worlds 2, a couple of months ago. Google is not launching into an empty field — it is defending a lead it established with Genie 3, and the interesting question is whether the lead survives once every lab can generate a world with sound.
For readers who follow the evaluation side rather than the demos, the useful frame is the one we used when WorldArena 2.0 put world models to the test on real robots: a world model is only as good as the consistency it holds under an agent's actions, and two minutes of coherence with no published eval is a claim, not a result.
Amazon has told its software engineers that token usage and code-generation volume are "deeply flawed" ways to measure whether AI is working. The memo, circulated internally in September and posted to the company's website on Tuesday, rewrites how Amazon defines the engineering role in the AI era and argues that teams should be judged on long-term outcomes instead — a position that lands awkwardly for every company that has spent two years counting adoption. "If your work does not benefit from current AI tooling, that is a finding, not a failing," the memo says, per Amazon's published version. CNBC reported the internal rollout and the company's response.
Amazon is not banning the metrics, just demoting them: token use and code volume should be read "with care and judgment" alongside other data. Marc Brooker, a vice president and distinguished engineer at AWS, told CNBC that the company now looks at feature velocity — how many launches and successful production changes a team makes — and treats the whole thing as a return-on-investment question, with an appetite for experimentation but no interest in spending tokens on problems that do not matter. An Amazon spokesperson said there are no caps on employee token use; the company shut down an internal usage leaderboard in May after discovering staff were tokenmaxxing to climb it.
What to watch next: whether Genie 4 reaches the Gemini API with a published price and an eval that measures coherence under agent control, and whether other large engineering organisations follow Amazon off the token leaderboard.
Should a world model that generates convincing audio and dialogue be watermarked everywhere, or is SynthID enough? Write to us via our contact page.
