On October 27, 2026, Google DeepMind announced Genie 4, an AI that generates ambient sound, music, and dialogue together with the world itself, as a research preview. Sound has been added to the "world model," which creates the next scene in response to a person's actions, and the announcement also described a setup in which Gemini watches the scene as it is generated and revises its plan. Genie 3 already offered real-time control and editing, so being controllable is not what is new this time. To evaluate this technology, which is still at the trusted-tester stage, as a training environment, you need to look beyond how natural the footage appears and check whether the same world holds together after repeated actions.
A generated world that now includes sound and dialogue
According to Google DeepMind's announcement, Genie 4 generates scenes that change in response to the user's actions, together with ambient sound, music, and dialogue. The explanation is that the world and its sound are generated simultaneously, rather than sound being added to finished video afterward. However, the statement that they are generated simultaneously does not tell us the model's internal architecture or how it was trained.
Users can control movement and the camera with a keyboard or similar input, and can change scenes with natural-language instructions. Movement is an operation that shifts the viewpoint within the world, while editing is an operation that asks for changes to the generated scene, so the two serve different roles. Genie 3 could already be controlled in real time and supported editing, so this is not a generational shift from "footage you watch" to "a world you can move through for the first time." What is new this time is the addition of sound and dialogue.
Google says that human motion, object collisions, and the depiction of water have also improved. These make the experience feel more realistic, but they should be read separately from any measurement of how accurately physical laws are reproduced. The published demonstrations are recordings of real-time control, but they carry a note saying scenes were shortened for explanatory purposes. The length of the footage cannot be taken as an upper limit on continuous operation or as a task success rate.
Gemini watches progress and corrects the plan
Google describes generating the world and deciding what to do within it as separate roles. Gemini Omni supports the realistic look and behavior, while a separate Gemini handles reasoning. Rather than thinking of Genie and Gemini together as a single autonomous robot, it is easier to understand them as one side that produces scenes and another that looks at the scenes and decides the next action.
The guide AI breaks a goal into multiple steps. It then observes the generated frames, judges how far things have progressed and whether a step succeeded, and revises the plan as the situation requires. Instead of handing over a long instruction at the start and waiting for the result, the setup feeds observations gathered along the way back into the next instruction. This division of labor lets it adjust its actions as the generated world changes. It is not a result from measuring improvements in task success rates.
The early robotics research on the official model page presents an example in which Gemini 4 Argon creates a multi-step plan and passes the steps as text to Genie 4, which keeps rendering. This is a demonstration inside a generated simulation. It does not show results from training an actual robot and transferring that control to a physical machine.
Even if a plan can be revised while watching the footage, whether object positions and states are preserved within that footage must be checked separately. That is why it matters what the comparison with the previous generation is actually measuring.
How does it differ from Genie 3's "minutes"?
Genie 3's "minutes" refers to environmental consistency, while Genie 4's "minutes" refers to continuous operation time, so they cannot be compared as improvement rates on the same metric.
Sorting the published items for both generations by what they describe makes the difference clear.
| Generation / source date | Published figure | What it describes | Comparison caveat |
|---|---|---|---|
| Genie 3 (Aug. 5, 2025) | Minutes | How long environmental consistency is maintained | Not the same metric as Genie 4's continuous operation time |
| Genie 4 (Oct. 27, 2026) | Minutes | How long it can be operated continuously | An improvement rate in consistency measured under the same conditions cannot be calculated |
| Genie 3 (Aug. 5, 2025) | 720p, 24 fps | Resolution and frame rate of generated video | Not a specification that carries over to Genie 4 |
| Genie 3 (Aug. 5, 2025) | Up to about 1 minute back | How far back it can visually remember | Distinct from continuous operation time and environmental consistency |
Table note: Compiled by matching each generation and what each expression refers to, based on Google DeepMind's Genie 3 announcement and the Genie 4 official model page. The latter reflects the October 27, 2026 research preview. These are not benchmarks run under identical conditions, and no figures were normalized or performance rankings assigned.
Therefore, this comparison supports neither "both are minutes, so there has been no progress" nor "it has become stable for longer than before." Applying Genie 3's image quality or frame rate to Genie 4 would mix in specifications from a different generation. In the official explanations we reviewed, a comparison of duration under the same conditions does not hold.
The Genie 3 announcement explained that errors tend to accumulate when the next frame is produced in sequence while referring to past generation history. This is a reason to evaluate separately whether footage looks natural and whether the world's state holds up over long operation. It is not evidence that Genie 4 uses the same internal mechanism, but it is a question that cannot be avoided if a generated world is to be used for training.
What the 2018 world model showed about training pitfalls
The idea of separating a model that predicts the world from a controller that chooses actions has precedent. The 2018 research by David Ha and Jürgen Schmidhuber, "World Models," separated a visual model that handles imagery, a memory model that handles the flow of time, and a controller that chooses actions. It combined them as V for vision, M for memory, and C for control, and trained a control policy inside a generative model.
That research presented an attempt to apply a control policy trained in a generated environment to the original game, VizDoom. The target was a game, not a real-world robot. Even so, there is value in a design that separately checks whether something works inside the learning world and whether it also works in the original environment that world imitated.
One of the problems the research showed was that the controller could exploit flaws in the generative model. Even if it seemed to have learned to solve the task, behavior that depends on the model's weaknesses will not work in the original environment. The better the results inside a generated world, the less guaranteed it is that they will be useful outside it.
This does not mean the same flaw has been found in Genie 4, nor does it mean the two use the same learning method or network. What past research offers is a question for evaluation. Even if there is a mechanism by which an AI observes a generated scene and judges it a success, that judgment and being able to accomplish the task in the original environment must be verified separately.
Availability and remaining limitations
What is offered at the time of the announcement is a research preview for trusted testers. Access for Google AI Ultra users through Project Genie, and for early-access partners of the Gemini API and Google AI Studio, is described as coming later. It is not an announcement that general availability has begun or that availability in Japan or pricing has been decided.
The limitations Google lists also include some that cannot be overlooked when evaluating it as a training environment. When operation continues for a long time, complex edits are added, or the user moves around, objects' appearance may change or they may fail to remain the same object. This is the problem of whether the same objects and states are kept before and after an action. Complex interactions with multiple independent agents remain a challenge, and there are limits to accurately reproducing real-world locations and to rendering legible text. Even with sound added, the accuracy of real-time speech remains an area of ongoing research.
Google explains that video and audio will carry SynthID, which identifies generated content. However, being identifiable does not guarantee the physical correctness of the world. Even if generated footage looks plausible and the observing AI accepts that a step succeeded, that does not mean real contact forces or positions were reproduced correctly.
For training uses, the judgment criteria are, first, whether an object's state holds up through repeated actions; second, whether success can be verified after an action; and third, whether results can be reproduced in the original environment or on a physical machine. The announcement shows no results from transfer to a physical machine. Once such verification is in place, a generated world with sound and dialogue could move from a place for creating experiences to a place for repeatedly testing behavior that also holds up outside it.