DIRTY STATIC / DISPATCH Reviews20 NOV 2026 • 14 MIN READ
A vast concrete portal frame in a dark brutalist hall opening onto a small grayscale explorable world of roads, trees and buildings, with a giant keyboard key and a caution yellow playhead above it and worn hazard stripes on the floor
ReviewsINTELLIGENCE FOR THE MACHINE AGE

Genie 4 API: 720p Worlds At 24 Frames

Google opened Genie 4 to developers on November 19, 2026, through the Gemini API and Google AI Studio, one day after it reached Project Genie subscribers. This research review checks the API as documented: the 720p stream at 24 frames a second, the native audio, the session limits, the quotas, and a price of about ten cents a second, and weighs who should build on it now.

Start reading ↓

Dirty Static // The Verdict

The first world model a developer can actually call, and a genuinely new kind of API: a steerable place instead of a finished clip. It earns a trial for short prototypes and agent tests. It does not earn a product plan yet, because a world lasts two minutes at most, drifts inside that window, and costs about $363 for an hour of play.

Best forDevelopers and creative teams who want to prototype an interactive scene, test a spatial idea, or run short agent and robotics simulations without building a game engine scene first

What works

  • A real developer route at last, through the Gemini API and AI Studio with a named model, printed prices and published quotas, one day after the consumer launch in Project Genie
  • Native audio arrives with the picture, covering ambient sound, effects and character dialogue that reacts while the session runs
  • Steering is direct and simple to grasp, with keyboard movement plus live text edits for weather, lighting, objects and events
  • Pricing is plain arithmetic, at 240 video tokens a frame and about $0.10 a second, so a full two minute session has a known ceiling of about $12.10

What does not

  • Sessions stop at 120 seconds, cannot be resumed, and leave no persistent world behind, so nothing built here carries into the next session
  • Google itself warns that detail and object permanence drift, with the strongest coherence put at the first 45 to 90 seconds and stability fading after 60 to 90 seconds
  • An hour of play costs about $363 at the printed rate, which rules out shipping this as the world inside a consumer game or app today
  • There are no published benchmark scores to check, only Google human and automated evaluations described in general terms, and no free tier to test with

Research review: based on published testing, specifications, and documentation, not hands on testing.

A video model hands you a clip. Genie 4 hands you a place, for two minutes, and then takes it away. On November 19 Google put its newest world model into the Gemini API as models/genie-4, a day after the same model reached Google AI Ultra subscribers inside Project Genie. Google calls it its first world model for developers to build with, and that claim checks out in the narrow sense that matters. Genie 3 never had a public API at all. This one has a model name, a price list, and quotas you can read before you write a line of code.

What you get for the money is a continuous 720p world at 24 frames a second, with sound made alongside the picture, that answers to a keyboard and to sentences typed while it runs. What you do not get is anything that lasts. Each session caps at 120 seconds, cannot be resumed, and leaves no scene behind to reopen. That tension, a real API wrapped around a deliberately brief and drifting world, is the whole review.

What It Is

A world model is a system that predicts how a place changes as you act inside it, then draws that change frame by frame. Genie 4 is Google DeepMind newest entry in that line. DeepMind announced it on October 27, 2026, as a research preview for trusted testers, and the model card followed on November 18. The consumer door opened the same week, inside Project Genie for eligible AI Ultra subscribers, with AI Pro members offered a waitlist place. The developer door opened on November 19, and it is the door this review covers.

The system is built in layers, and Google is unusually open about the wiring. The model card says Genie 4 is adapted from Gemini Omni Flash and built on top of it, with Gemini models supplying reasoning and steering in real time. In the October announcement, DeepMind named the division of labour. Gemini Omni anchors how a simulation looks and behaves. Gemini 3.8 Flash rides inside the loop as a guiding agent, watching the frames, breaking a goal into steps, and checking that each step landed. Genie itself renders the immediate physical reality, pixel by pixel, sound included.

That sound is the headline change from Genie 3. Genie 4 is DeepMind first world model to generate audio natively, in the same pass as the picture rather than as a track laid on afterwards. Ambient sound belongs to the setting, wind in trees or waves at a beach in Google examples, and characters hold dialogue that reacts to what you do and to events around them. DeepMind also claims a step forward in physics and in people, with aerodynamics and collisions simulated and characters that move more naturally than in earlier versions. Those are company claims about quality. The model card says human and automated evaluations found clear gains over Genie 3 in photorealism, fluid movement, and steering, but Google has published no scores, no benchmark table, and no third party results for anyone to check.

Genie 4 on the Gemini API, checked November 20, 2026

Model
models/genie-4, through the v1beta Interactions API and Google AI Studio
Output
720p video at 24 frames a second, fixed 16:9, with synchronized native stereo audio at 48 kHz
Inputs
A text prompt, a single reference image, or both, sent as one world setup event
Controls
WASD and arrow keys for movement and camera, live text prompts for scene changes and dialogue
Transport
WebRTC for browser apps, WebSocket for server code, with single session ephemeral tokens
Session limit
120 seconds maximum, no resume after a session ends
Input price
$1.50 per million tokens for text and reference image input
Text output price
$9 per million tokens
Video output price
$17.50 per million tokens, audio included, at 240 tokens a frame
Effective rate
About $0.10 a second, $6.05 a minute, $12.10 for a full session
Not supported
Free tier, Batch API, Flex inference, Priority inference, live microphone or webcam input, function calling, Search grounding, structured outputs

Real World Experience

This is a research review. Dirty Static did not run Genie 4 sessions, and nothing here is firsthand testing. The assessment draws on the Google announcement of the API release, the DeepMind announcement and model card, the Gemini API documentation as summarised by Unite.AI and TestingCatalog, and independent price arithmetic from AI Press Box, all checked on November 20, 2026.

Start with the feel of the thing, as the documentation describes it. A session opens with one setup event holding your prompt, your image, or both, and the guide puts the wait for the first frame at about five to nine seconds. After that, keyboard actions land fast. The implementation guide reports a visual response of about 0.2 to 0.5 seconds for keyboard moves at low network latency, while a text change, a new storm, a new character, a line of dialogue, takes roughly one to two seconds to show. Network latency above 150 milliseconds noticeably dulls the steering, which is worth knowing before anyone blames the model for a sluggish demo run over a distant connection. One detail in the coverage deserves a raised eyebrow. Unite.AI reports sessions served only from data centres in the United States, while TestingCatalog reports serving from the United States, Western Europe, and Japan. The summaries disagree, so check the region list in the current documentation before you design around latency.

A concrete control desk with chunky levers and a WASD keyboard cluster overlooking a grayscale miniature landscape of hills, a river and a small town, with one cloud picked out in caution yellow
Steering is the product here. Keyboard movement answers in about half a second on a good connection, while a typed change to the scene takes one to two seconds to appear. Illustration: Dirty Static house style

Then there is the clock, and the clock is strict. The world is generated over a rolling window in time rather than held as a permanent 3D scene. Google documentation, as reported by Unite.AI, warns that looking away from an object for eight seconds or more, or turning a full circle, can move or reset landmarks, doors, and buttons. The model card says the same thing in politer language. Detail and object permanence can shift or drift as a run lengthens, after several complex prompt edits, or while navigating. Google puts the strongest stretch for coherence and permanence at the first 45 to 90 seconds, and warns that visual stability may degrade after 60 to 90 seconds. Read those two numbers together and the honest summary is this. You get about a minute of world you can trust, inside a two minute allowance you are paying for.

Billing has one trap, and it is a simple one. Closing a browser tab detaches the transport without stopping generation on the server, so the meter keeps running. The documentation tells developers to call complete() to end a session cleanly, or cancel() to abort it, and to treat that call as part of cost control. The saving grace is the cap itself. A session nobody stops can cost at most about $12.10, because it cannot run past 120 seconds. AI Press Box did the wider arithmetic from the pricing page. An hour of Genie 4 play costs $362.88, spread across at least 30 separate worlds, and that hour costs about 45 percent more per second than Google own clip generator at the same resolution, Gemini Omni 1.2 Flash, which makes fixed clips of up to 15 seconds. Our tutorial on making a short video with Gemini Omni 1.2 Flash shows the clip route. Genie charges more because the world answers back, and whether that premium is worth paying depends entirely on whether your idea needs an answer.

Cost of one hour of generated 720p video, dollars (pricing page figures, November 20, 2026)

Genie 4, interactive world$362.88
Gemini Omni 1.2 Flash, fixed clips$250.21

Genie 4 is interactive and streamed in sessions of up to 120 seconds. Gemini Omni 1.2 Flash makes fixed clips of up to 15 seconds and is not interactive. Rates worked out by AI Press Box from the Gemini API pricing page: Genie 4 at $0.1008 a second, Omni 1.2 Flash at $0.0695 a second.

Quotas keep early use small by design. A Tier 1 paid project gets 2 concurrent sessions and 50 session requests a day. Tier 2 gets 5 concurrent and 250 a day. Tier 3 gets 10 concurrent and 1,000 a day. Control messages inside a session do not eat the daily count, which is sensible. Even so, ten players at once is the ceiling at the top tier, and that number alone tells you what Google thinks this launch is for. Ten people can test a prototype together. Ten thousand cannot play a shipped game.

What Works

The access itself is the first strength, and it should not be undersold. World models have lived in demos, waitlists, and subscriber playgrounds. Genie 3 spent its whole life that way. Genie 4 gives a developer a documented session flow, standard transports, ephemeral tokens that keep the main API key off the client, and starter apps in AI Studio for a first look. A team can go from reading the documentation to steering a world inside its own app, and that has simply not been true of this model family before.

The audio is the second strength, because it arrives as part of the world rather than as decoration. A scene that cracks, splashes, and talks back while you steer it is a different prototype from a silent one. For previz work, for dialogue tests, and for agent training where sound carries information, generating the sound in the same pass removes a whole stitching step. DeepMind lists audio accuracy as an open problem, and fairly, but the direction is right and no rival in the API comparison below ships reactive dialogue as a standard part of the stream.

The third strength is how legible the deal is. The price is one multiplication a developer can do on paper. The session cap is a hard number. The quotas are a small table. The limits, drift, weak text rendering, shaky spatial accuracy, are printed in Google own model card rather than discovered by users. A launch that tells you its world forgets things after eight seconds of looking away is a launch you can plan around, even when the plan is to wait.

What Does Not

Persistence is the big absence, and it shapes everything else. There is no scene to save, no world to reopen, no session to resume, and no export of the place you just walked through. Every new session starts from a prompt and an image again. For prototyping and for short agent tests, that is acceptable, even tidy. For anything a user is meant to return to, a level, a showroom, a training course with a second lesson, the API offers nothing to build on today. The world is a performance, not a place.

A long concrete film strip corridor unrolling through a dark hall and dissolving into static past a caution yellow marker gate, with a small figure walking inside it
A Genie 4 session is a strip of time, not a stored scene. When the two minutes end, the world ends with them, and the next session starts again from a prompt. Illustration: Dirty Static house style

The price is the second problem, and it is a problem of shape rather than size. Ten cents a second is a fine price for a test and a brutal one for a product. At $6.05 a minute, a consumer experience that keeps a player inside a world for an evening would burn through hundreds of dollars of generation per player. Google sells the same model to consumers a different way, inside Project Genie on an AI Ultra plan starting at $99.99 a month with daily limits and no per world credit charge, per the Project Genie help pages. That split is the tell. The subscription is where Google expects people to play. The API is where it expects developers to sample.

The evidence gap is the third problem. Every quality claim in this launch, the realism, the physics, the natural movement of people, the dialogue, rests on Google demonstrations and Google evaluations. The model card describes the evaluation approach and reports clear improvement over Genie 3 without publishing a single score. Trusted testers quoted in the October announcement were picked by Google and spoke about draft mode potential, which is encouraging and is not verification. Until outside teams publish their own sessions and failure rates, a buyer is pricing a promise.

Competition

RouteWhat you getCost shapeThe trade
Genie 4 APIA steerable 720p world with native audio, streamed live, gone when the session endsAbout $0.10 a second, $12.10 per full session, paid tier onlyThe only DeepMind world model with a developer API, and the briefest world on this list
Project GenieThe same Genie 4 model in Google own browser experience, with an agent and a community galleryInside an AI Ultra plan from $99.99 a month, daily limits, no per world credit chargeThe cheapest way to play for hours, with no way to put the world inside your own product
Gemini Omni 1.2 FlashFixed video clips at the same 720p, up to 15 seconds, no steeringAbout $0.07 a second of finished clipCheaper per second and simpler, for any idea that does not need the world to answer back
World Labs MarbleA persistent 3D world you can revisit, edit, expand, and export as splats or meshesAbout $1.20 a world through its API, with app plans from free to $95 a monthA world file that stays put, generated rather than streamed live, and silent where Genie talks

The sharpest contrast is with Marble. Marble hands you scene data. You can leave, come back tomorrow, walk a different path, and export the place into an engine. Genie 4 hands you two minutes of behaviour, including sound and people who respond, and keeps nothing. Those are different products wearing the same world model label. Pick by what your idea cannot live without. If it needs a place that persists, Genie is not your tool yet, whatever the demo does to you. If it needs a place that reacts, with voices, inside the next two minutes, nothing else on the list does that through an API today. Other streaming rivals exist, including Decart Oasis 3 at a reported $0.02 a second, but their published detail is thinner than Google documentation for this launch, so they sit outside the table rather than inside it.

Final Verdict

Build a prototype on it this week if your work touches interactive scenes, spatial design, or agent training. The cost of finding out what a steerable world does to your idea is about twelve dollars a session at the very worst, the steering model is easy to grasp, and the audio gives even a rough test a life that silent worlds lack. Keep the test inside the first minute of each session, where Google says the world holds together best, and write down what drifts.

Do not build a product on it yet. A two minute cap with no resume, no saved scene, ten concurrent sessions at the top tier, and $363 an hour of play add up to a sampling tool, not a foundation. The right posture is the one Google own pricing suggests. Sample through the API, play through Project Genie if you hold an Ultra plan, and revisit the question when a Genie model offers a world that survives the end of its session. The door opening matters more than the room behind it this time, and the door is genuinely open.

Sources

■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘