Models#Video generation#Model inference
Google DeepMind unveils Genie 4, a world model that generates audio and rewrites the world mid-session
Genie 4, DeepMind's new real-time world model: native audio, mid-session text edits, Gemini 3.8 Flash in the loop — research preview now, Project Genie next.
Google DeepMind released Genie 4, a real-time world model, on October 27. Visuals, music, ambient sound and character dialogue come out of the same model in real time — DeepMind calls it “our first real-time world model to generate audio natively” — and you can rewrite a running session with one natural-language line: change the weather, drop in new people, move the camera, no re-generation needed. It opened as a research preview to trusted testers on launch day and reaches Google AI Ultra subscribers through Project Genie next. The blog post itself quotes no frame rate, resolution or session-length numbers.
Key points
- Audio is generated by the world model itself for the first time: music, ambience and dialogue that shift with the story, in a model that runs interactively.
- Session consistency roughly doubled to about 2 minutes — a press figure from The Verge and TechCrunch; the official blog states no length at all. Genie 3 managed about 1 minute.
- Ordinary users still can’t get in: trusted testers first, then AI Ultra (from $100/month per Ars Technica), with Gemini API and AI Studio pricing unannounced.
Background
DeepMind’s previous generation stopped halfway. Genie 3, unveiled in August 2025, knocked gaming stocks down on arrival — and per Ars Technica’s retrospective, never expanded beyond the Project Genie web experience. A year later the lane is crowded: in TechCrunch’s accounting, Runway shipped GWM Worlds 2 two months ago, AMD is acquiring World Labs, Amazon is reportedly in talks for Decart. The research side is moving too — we covered SkeLAM, which trains an action-conditioned world model from about 10 hours of robot data, and Yann LeCun’s bet that world models, not bigger LLMs, are the next breakthrough. What Google demos under the “world model” label is not what LeCun means by the term, a distinction one HN commenter presses hard.
The facts
The core configuration comes from the launch blog (October 27, signed by Raia Hadsell, Diego Rivas and Hang Qi); duration and pricing are marked where they rest on press reporting:
| Dimension | Claim | Source |
|---|---|---|
| Audio generation | “First real-time world model to generate audio natively” | Official blog |
| Mid-session edits | Weather, new people, camera moves — one line, live | Official blog |
| Frame-by-frame steering | Gemini 3.8 Flash in the loop: splits goals, verifies steps, re-plans as the world shifts | Official blog |
| Realism anchoring | Gemini Omni anchors appearance and behavior | Official blog |
| Session consistency | ~2 minutes vs ~1 for Genie 3 | The Verge, TechCrunch (no official number) |
| Pricing | AI Ultra from $100/month; API and AI Studio unpriced | Ars Technica |
| Watermarking | SynthID on all video and audio output | Official blog |
| Robotics demo | “Gemini 4 Argon and Genie 4” drive Apptronik’s Apollo humanoid playing checkers | Official blog |
Dialogue gets its own billing: “Genie 4 introduces interactive dialogue that adapts on the fly” — the saxophonist on the lake responds to your instructions, and lines change as the plot moves. The same post lists “complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency” as active research areas, and concedes that “visual details and object permanence can still shift or drift over time”, especially in long sessions.
What others say
The Verge’s Jay Peters relays Google’s three pitch uses — film pre-visualization, game prototyping, robot training — but his hands-on memory of the last generation is sour: he used Project Genie to “generate bad Nintendo knockoffs” last year, and couldn’t try Genie 4 this time. Ars Technica’s Ryan Whitwam is most skeptical about speech: “The biggest leap, if it works, is the addition of human speech” — two doubts packed into one sentence. TechCrunch’s Rebecca Bellan quotes DeepMind engineer Diego Rivas placing the value “in the ideation phase, in the storyboarding phase,” and adds the competitive frame.
The Hacker News thread (200+ points, 150+ comments) is most useful where it frame-checks the demos: in the dragon-flight clip, between 1:25 and 1:29, the right side of the castle — catapult included — vanishes when the camera swings back; in the dressing-room scene, the walking man wears a grey suit on his left half and a grey dress on his right. On the replace-game-engines narrative, practitioners push back with input-latency and compute-cost math, and one points out the game industry still hasn’t digested plain LLMs without catastrophic edge cases. Another comment worth keeping: Google is accused of redefining “world model” — shrinking LeCun’s “a system that understands physical causality” into “text-to-playable-video.”
Our take
The genuinely new thing in Genie 4 is not realism, it’s where control sits: the model graduates from “a thing that renders clips” to “a director’s console over a running world.” For storyboarding and pre-vis workflows that shape is immediately usable — Rivas’s own framing points there. For anyone hoping to build a game on it, the limitation section is the answer: a ~2-minute consistency window doesn’t carry a shipped product, so wait.
Note the familiar flagship rhythm, too. Genie 4 follows the same path as Gemini 4 Argon on September 30: trusted testers, then subscription tiers, no API pricing — while the previous generation never left the web toy. Before you join the waitlist, ask what’s different this time; HN’s kouteiheika puts the counter-case bluntly: zero technical details released, not even a parameter count.
How to try it
- Web: the Project Genie waitlist is open; AI Ultra runs from $100/month per Ars Technica.
- Developers: the Gemini API and Google AI Studio early-access program is rolling out in the coming weeks, pricing unannounced.
- No public evals or third-party benchmarks yet — between joining the waitlist and using it reliably sits the very limitation list DeepMind published.
What to watch
Four checkable next steps: the AI Ultra rollout timeline, a Gemini API price, third-party tests against the 2-minute consistency window, and the hybrid direction the blog teases — plugging Gemini’s planning into Genie’s pixel-level generation — actually shipping.