Google's Genie 4 Adds Sound, Speech and 2-Minute AI Worlds After Game Stock Rout

Google DeepMind on Tuesday, October 27, 2026, unveiled Genie 4, a new "world model" that generates explorable, photorealistic scenes in real time, now with its own sound, speech and more lifelike people. It is a research preview for a small group of trusted testers for now, and Google said it will reach Google AI Ultra subscribers and early developer partners in the coming weeks.

The launch matters beyond the lab. When Google opened up the previous model to paying subscribers in January, video game stocks had one of their worst days of the year, with engine maker Unity closing down 24%. Genie 4 roughly doubles how long a generated world stays consistent and adds features that make its output look much more like a playable scene, so investors, game studios and creators are all paying attention.

Key facts at a glance

• What it is: Genie 4 is Google DeepMind's fourth-generation world model, software that builds an interactive environment frame by frame from a text or image prompt instead of from pre-made game assets.
• What is new: native audio, including music, ambient sound and interactive dialogue; better physics such as collisions and aerodynamics; more natural human characters; and live text edits mid-scene.
• How long worlds hold together: up to about two minutes of consistent scene, double the limit of Genie 3, Google said in its briefing.
• Who can use it: trusted testers today, then Google AI Ultra subscribers through the Project Genie app and early access partners of the Gemini API and Google AI Studio. Ultra plans start at $100 a month, and developer pricing has not been announced.
• Safety labelling: all video and audio Genie 4 generates will carry SynthID, Google's invisible watermark for AI-made content.

How we got here

Genie began as a research project. Genie 1 arrived in February 2024, Genie 2 followed in December 2024, and Genie 3 was shown as a research preview in August 2025. None of those were products ordinary people could try.

That changed on January 29, 2026, when Google launched Project Genie, an experimental web tool powered by Genie 3 for Google AI Ultra subscribers in the US. Sessions were capped at about one minute, ran at 24 frames per second and topped out at 720p resolution. Generated worlds could also shift mid-session, with one demo showing a race track changing during play.

The market reaction the next day was sharp. On January 30, Unity closed down 24%, its worst one-day fall since 2022, while Roblox dropped 13%, CD Projekt 8.7%, Take-Two Interactive 7.9% and AppLovin 17%. Investors appeared to worry that prompting a world into existence could one day undercut the engines, tools and studios that build games today.

Google has kept moving since then. On October 7 it announced Playground, an AI game creation tool built with Unity, which lifted Unity shares about 4% that day while Roblox fell about 4%.

What Genie 4 actually adds

The biggest change is sound. Genie 4 is the first DeepMind world model to generate audio natively, so a beach scene comes with crashing waves and a forest with wind in the trees. Users can also describe the audio they want in the prompt.

Google said the "true frontier" is human speech. Characters in a Genie 4 world can hold conversations and react to what the user does, with dialogue that adapts as the scene changes. That is a large step from the silent worlds of Genie 3.

The second change is realism. Google said Genie 4 better simulates physics such as aerodynamics and collisions, and that people in its scenes move more naturally and respond to their surroundings. Its demos include a watermelon on a stone plinth, a maple seed falling into grass and a crowded 1980s Japanese train station.

The third change is real-time steering. While a world is running, a user can type a request such as changing the weather, moving the camera or adding a new person or object. Navigation still uses standard keyboard controls, and Gemini adapts each text request to fit the surrounding scene.

Here is how the two public generations compare, based on what Google has disclosed:

Gemini as the director

Genie 4 leans heavily on Google's Gemini models. Google said Gemini Omni anchors scenes in real-world appearance, while Gemini 3.8 Flash acts as a guiding agent in the loop. As Genie renders the world, Gemini watches each frame, breaks a goal into steps, checks that each step happened and changes its plan if the user steps in.

Tejas Kulkarni, a research director at Google, described where this is heading. He said the future lies in hybrid world models, where Gemini writes code and text and Genie takes that in real time to simulate reality at different levels of detail.

Google also showed a robotics experiment. Its Gemini 4 Argon model broke a task into steps that Genie 4 then simulated, in this case Apptronik's Apollo humanoid robot playing checkers. The idea is to test robot behaviour in a simulated world before trying it in the real one, which Google describes as a key use for world models alongside games and creative work.

Who is testing it, and the competition

Google said trusted testers come from film, journalism, education, gaming and engineering. Filmmakers are using it as a "draft mode" to run several takes of a scene and adjust camera angles, lighting and dialogue on the fly. Diego Rivas, a Google DeepMind engineer, said the real-time element is most valuable in early ideation and storyboarding.

Photojournalists are testing how a single archival photo can become an explorable scene, for example to show how a flash flood or wildfire unfolds. Game developers are using it to go from an idea to an explorable prototype in seconds, testing visual styles and mechanics before writing engine code.

Google is not alone. AI video startup Runway released its second world model, GWM Worlds 2, a couple of months ago. The field is also consolidating: AMD is acquiring Fei-Fei Li's World Labs, Amazon is in talks to buy Decart, and carmaker Rivian has bought Third Dimension to bring simulation in-house for self-driving testing.

The other side: limits and skeptics

Google itself is clear that Genie 4 is not a finished product. The company said visual details and object permanence can still drift over time, especially in longer sessions, after many prompt edits or when moving around a scene. It also listed complex physical interactions, accurate real-time audio, legible in-world text and spatial accuracy as unsolved problems.

Rivas stressed during the briefing that it is still early and that the end-to-end product experience is not yet clear. Two minutes of consistent scene is a big gain, but it is a long way from the hours of stable, repeatable play a commercial game needs.

Some analysts thought January's sell-off went too far. Analysts at Polish broker mBank called the drop in game stocks "unjustified," pointing to short sessions, limited controls and no reliable persistent gameplay. Unity and Roblox shares rebounded in early trading the following Monday.

What it means for you

If you play games, nothing changes yet. Genie 4 makes short, explorable scenes, not full games with saves, rules and multiplayer. The near-term impact is more likely to show up in how studios prototype ideas than in what you buy.

If you create video or games, the most practical use today is rapid previsualisation: trying a scene, a look or a camera move before committing a team to build it. Access will be limited and costly at first, since Ultra plans start at $100 a month in the US.

If you invest in game stocks, January showed how quickly a demo can move prices. Roblox reports third-quarter results on October 29, and executives on that and other earnings calls this week may be asked about AI world models.

If you care about deepfakes, Google said everything Genie 4 generates will carry SynthID watermarks in both video and audio. That helps platforms detect AI-made content, but it only works where detection tools are used.

What happens next

Google said Genie 4 will roll out to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio in the coming weeks, with a waitlist open now. It has given no exact date and no developer pricing.

Open questions include whether access widens to cheaper AI Pro and AI Plus plans, whether sessions grow past two minutes, and how well the dialogue and physics hold up outside curated demos. Google said its longer-term goal is hybrid world models that combine Gemini's reasoning with Genie's pixel-level simulation.

The bottom line

Genie 4 is a real step up from the model that shook game stocks in January: it adds sound and speech, more lifelike people and physics, and twice the stable session length. It is still a research preview with clear gaps that Google itself lists.

Practical takeaways:

1. Join the Project Genie waitlist if you want to try it early, and expect Ultra-level pricing at first.
2. Treat it as a prototyping tool, not a game engine replacement, for at least this first rollout.
3. Watch the October 29 Roblox results and other game company calls for how executives respond.
4. Check for SynthID labels when you see unusually realistic AI scenes, and expect platforms to use them more.