Genie 4
A world that follows
Google says its Genie 4 draws a world you can walk through, hear and change by typing, and its own pages give no figures for how well that world holds up.
Robots · Checked 13 Nov 2026 · 11 sources · 9 min

Nobody is surprised to find a cup where they left it. You set it on the kitchen table, look out of the window, look back, and there it is. A cup does not need to be watched in order to stay put.
A video game has to arrange the same thing on purpose. Behind the picture, the game holds a record of its world, something like a list: a table here, a cup on it there. Many times a second, a program works out what you would see from where you stand, and draws it. Turn away and the cup leaves the picture but stays on the list. Turn back and it is drawn again from the same entry, in the same place. The price is labour. Before anyone can walk in, somebody has to build the table and the cup and write the rules, such as the one that says a cup pushed off an edge falls.
A newer kind of program skips the building. You type a description of a scene, and the program produces a moving picture of it, then keeps changing that picture in answer to what you do, as if you were walking through a place. The field calls such a program a world model, and calls the typed description a prompt. On 27 October 2026 Google DeepMind, Google's AI research arm, announced a world model named Genie 4, released as a research preview to trusted testers.
By Google's account, sound comes with the picture, and people in the demos move through the world with keyboard and mouse. Neither the announcement nor the Genie product page states how large the model is, how sharp its picture is, how many pictures it draws each second, what it costs or how it scores on any standard test.
None of that says where the cup goes when you look away. Suppose the scene you asked for is the kitchen, cup and all. A moving picture is a run of still pictures shown fast enough to pass for motion, and each still is called a frame. So the puzzle of the cup comes down to one frame, the first after you turn back toward the table, and to what the program has to go on when it draws it.
Google has described this for the model before Genie 4. Genie 3, announced on 5 August 2025, draws its frames one at a time, each depending on what it has generated so far, so that its own output is fed back in as input. The field's name for this is auto-regressive generation. Google's Genie 4 pages do not say whether Genie 4 does the same.
By that description, the earlier frames have to do the job the list did in the game. While you look out of the window, nothing on screen shows the cup. When you turn back, a cup appears on the table only if the model, working from frames in which a cup was visible a few seconds earlier, draws one there again. Keeping things in existence while they are out of sight is called object permanence. The game got it from one entry on a list. A model of this kind has to manage it afresh each time the table comes back into view.
Such an arrangement has a weakness built in. Each frame is drawn from the frames before it, mistakes included. If one frame puts the cup's handle on the wrong side, the next has no list to check against. It finds a cup with its handle on that side and carries on, so an error can become part of the record every later frame is drawn from. And a world that answers each keypress has to make its frames as fast as it shows them, which is what the field means by running in real time.
So the thing to ask of such a model is how long it holds together. The research version of Genie 3 ran at 24 frames per second, a rate that cannot be assumed for Genie 4, and by Google's account it stayed consistent for a few minutes. One minute at that rate is 24 times 60, or 1,440 frames. Suppose a model slipped on one frame in every 1,000. With 1,440 frames in a minute, that is more than one slip a minute, each handed on to the frames that follow.
Google describes trouble of this kind in Genie 4, which it calls an early, experimental model whose visual details and object permanence can drift, especially in longer runs, after multiple complex prompt edits or while navigating. Several of the limitations Google lists for Genie 4 repeat, nearly word for word, the list it published for Genie 3.
The game got a second thing from its bookkeeping, a cup that behaves. Pushed off the table it falls, because somebody wrote a rule for falling. Programs that generate pictures generally get no such rule: they are trained on examples and produce what the examples make likely to come next, which most of the time looks right. Google lists complex physical interactions, and the accuracy of real-time audio, among the things still under active research in Genie 4, and The Batch reported a physics slip in one of Google's own clips, where a marble bust dropped onto a watermelon settles on its plinth instead of toppling over the side.
Two minutes, as told to reporters
Google's own pages put no figure on how long a Genie 4 world lasts. The product page says only "a few minutes of continuous interaction, rather than extended sessions". The number that went round came from a press briefing: Bloomberg reported that Google said simulations can run up to two minutes, roughly double Genie 3, and TechCrunch gave the same figure. Hang Qi, a research engineer at Google DeepMind, told Bloomberg that reaching the limit does not necessarily mean a scene stays fully consistent.
The doubling is hard to read, because two lengths are on record for Genie 3. One is the few minutes of its research version. The other belongs to Project Genie, the consumer prototype built on it, which launched on 29 January 2026 with generations limited to 60 seconds. Two minutes is double that 60-second cap, and no source says which limit was doubled.
Lengths on record, each from its own source
Three releases and the length each was given
5 August 2025
Genie 3 is announced. Its research version runs for a few minutes
Google DeepMind29 January 2026
Project Genie, the consumer prototype built on Genie 3, launches with generations limited to 60 seconds
Google27 October 2026
Genie 4 is announced. Google tells reporters its simulations can run up to two minutes
Bloomberg
Every demonstration of Genie 4 is a recording Google made, and a recording by the maker is weaker evidence than it looks. Whoever makes a recording chooses which session to keep, how long to let it run and where to cut. A clip in which the cup is still on the table after the view swings back shows that this can happen. It cannot show how often. The product page marks its demos "Sequences shortened". The Batch noted that Genie 4 has no technical report and that Google offered no quantitative evaluations, meaning results with numbers attached. No independent hands-on test was found.
An outside test needs outside access, and access is narrow. Beyond the trusted testers, Google says Genie 4 is "coming soon" to Google AI Ultra subscribers in Project Genie and to early access partners through its tools for software developers, and others can join a waitlist.
A second model watching the frames
By Google's account, a user can type an instruction mid-simulation, such as changing the weather, moving the camera or adding a person or object, and Gemini, Google's family of AI models, adapts the prompt to the scene. Google describes Genie 4 working with separate Gemini models, some of which supply the reasoning that keeps a simulation coherent. While a session runs, Gemini 3.8 Flash acts as "a guiding agent in the loop", an agent being a program that works toward a goal on its own. It observes the scene, breaks a goal into steps and reviews incoming frames to check that each step succeeded, revising its plan when the user intervenes. Google has published no description of the architecture or the training data, so how these parts are joined is not public.
The checking is where one outside publication saw a problem. Push the cup toward the edge of the game's table, and whether it falls is settled by the list and the rules, whatever you were hoping for. In one experiment Google describes, a Gemini model broke a task into steps that were fed to Genie 4 as text captions while it drew a humanoid robot, Apptronik's Apollo, playing checkers. There the plan was part of what the frames were drawn from.
The Batch argued that a planner checking each step's success can drift the world toward the intended outcome. Suppose the goal in the kitchen is for a robot hand to pick up the cup. The worry is that the cup ends up in the hand because the plan called for it, whether or not a hand closing at that angle would have held a real cup. Google claims "reliable causality", which is to say dependable cause and effect, for its Gemini loop. Neither side offers a measurement.
Somewhere for robots to practise
What rides on that disagreement depends on what Genie 4 is for. The chief executive of Alphabet, Google's parent company, in earnings remarks on 28 October 2026, called Genie 4 a building block for training robots and agents, a purpose the DeepMind post also states. Robots often practise a task in a simulated world first, because in the physical world every try takes time and can break things. A hand-built simulator carries the game's cost, since every kitchen has to be built. If a typed sentence can supply the kitchen, that cost falls away, but the practice is worth having only if the generated kitchen fails the robot where a real one would. AI Business describes this as the reality gap, whether performance in simulation transfers to the physical world.
The checkers clip is a rendering, which is a picture of a robot drawn by the model. The Batch observed that Google described no robot control system in the loop. No statement from Apptronik was found, and no source shows the software that steers a robot being trained or evaluated in Genie 4.
In AI Business, Manish Jain of Strategic Horizon Research said enterprises would need evidence that generated conditions match the real task. Evidence of that kind would be a count, made by trying the same task many times in the generated kitchen and in a real one and seeing how often the two end the same way. No source reports such a count for Genie 4.
Check it yourself
- Open the Genie 4 waitlist form and read its title to confirm that access is by application.
- Open the Genie 4 announcement and read the availability paragraph to confirm the research preview for trusted testers and the "coming soon" route.
- Open the Genie product page and read the limitations list to confirm the words "a few minutes of continuous interaction" and the label "Sequences shortened" on the demos.
- Open Project Genie on Google Labs and search the page for Genie 4 to see whether it has reached the consumer product.
- Read the Genie 3 announcement for its account of generating frames auto-regressively, its 24 frames per second and its limitations list, then compare that list with the one for Genie 4.
- Read Bloomberg's launch report for the two-minute figure and Hang Qi's caveat, and confirm that the figure comes from the press briefing.
- Read The Batch's analysis of the watermelon clip, then play the clip on the DeepMind post under "Pushing the frontier in realism for physics and human action" to test the reported physics slip against Google's own recording.
Sources
- Genie 4: A new generation of world simulation. Google DeepMind, 27 October 2026.
- Genie, model page. Google DeepMind.
- Genie 3 announcement. Google DeepMind, 5 August 2025.
- Analysis of Genie 4 and other world models. The Batch, DeepLearning.AI, 6 November 2026.
- Report on Genie 4, by Davey Alba. Bloomberg, 27 October 2026.
- Report on Genie 4, by Rebecca Bellan. TechCrunch, 27 October 2026.
- Project Genie launch announcement. Google, The Keyword.
- Alphabet Q3 2026 earnings call, CEO remarks. Google, The Keyword, 28 October 2026.
- Report on Genie 4 and robot testing, by Kinza Yasar. AI Business, 2 November 2026.
- Apply for Genie 4 access in Project Genie. Google Forms.
- Project Genie. Google Labs.