Genie 4 Release

Google DeepMind announced Genie 4 on Oct 27, 2026, a world model that generates explorable scenes in real time with their own sound and changes them on a typed prompt mid-run. As of Nov 8, 2026 it is a research preview for trusted testers, with Google AI Ultra and the Gemini API promised as "coming soon".

0:16 Google DeepMind announced Genie 4 on Oct 27, 2026, and as of Nov 8, 2026 almost nobody can use it: it is a "research preview for trusted testers", with a waitlist and access for Google AI Ultra subscribers and Gemini API partners promised as "coming soon", with no date. Genie 4 is a world model: from a typed description or a single image it generates a scene you walk through with the keyboard, frame by frame in real time, and it keeps working out what happens next. In the clip from Google DeepMind's launch film, a still image becomes a place to explore with its own ambient sound, the camera moves, and the scene changes while it runs.

What Genie 4 adds over Genie 3

DeepMind's announcement names four changes:

Sound. Genie 4 generates audio natively: music, ambient noise and dialogue with characters inside the world. Genie 3's worlds were silent.
Steering mid-run. A typed prompt changes the running scene, such as the weather, the camera or a person or object added, without starting over.
Physics. It simulates aerodynamics and collisions, which DeepMind presents as the basis for training robots in simulation. One experiment paired Gemini 4 Argon and Gemini Robotics with Genie 4 to simulate an Apptronik Apollo humanoid playing checkers.
Other models in the loop. Gemini Omni anchors how things look, and Gemini 3.8 Flash acts as a guiding agent that interprets what the player does.

DeepMind also says what does not work yet: visual detail and object permanence drift over longer sessions. Every output carries a SynthID watermark.

6:36 The difference sound makes is clearest in Google's own Release Notes episode, where DeepMind's Diego Rivas starts a scene of a band muted, prompts it to life and then unmutes it: the band is playing, and he calls audio "the obvious thing that you didn't know you were missing". He compares the characters with a Genie 3 baseline of the same frame, which looked "gamified", while Genie 4's look photographic.

Who can use Genie 4, and what it costs

Nobody can buy Genie 4 today. Google publishes no price for it, and it is on neither the Gemini API models list nor the Gemini API pricing page (both read Nov 8, 2026). The consumer route will be Project Genie, a feature of the Google AI Ultra plan at $99.99 or $199.99 a month in the US; on Nov 8, 2026 that page still described Project Genie as running "Genie 3", so the switch to Genie 4 has not reached it. Google has stated no regions, session length, resolution or rate limits.

Announced	Oct 27, 2026
Access now	Trusted testers, by waitlist
Coming	Google AI Ultra (Project Genie), Gemini API and Google AI Studio early access partners, no date
Price	Not published; Project Genie needs AI Ultra, $99.99 or $199.99 a month
Watermark	SynthID on every output
Limits	Context, session length, resolution and regions not stated

What we think

We think Genie 4 reaches Google AI Ultra subscribers in Project Genie before it reaches the Gemini API. Google already sells Project Genie, on Genie 3, inside Ultra, so the consumer route only swaps the model, while an API needs per-second prices and rate limits that Google has not published.

Confidence: Likely. Check by: Mar 2027.

We expect the first Genie 4 product that anyone can use to be capped at short sessions, about a minute or less, because DeepMind itself says detail and object permanence drift as a session runs long.

Confidence: Possible. Check by: Jun 2027.