Google Genie 4 Gemini API Launch: Build Playable 720p Worlds for About $0.10 a Second

Google has handed developers the keys to its most unusual AI model yet. The Google Genie 4 Gemini API is now live: anyone with a paid Gemini API key can build interactive worlds that generate themselves in real time — with sound, moving characters and dialogue — instead of watching a pre-rendered video.

Genie 4 is not another chatbot. It is what Google DeepMind calls a world model: type a description, or upload a reference image, and it creates a place you can walk through and change with new prompts while the simulation runs. It arrived in the Gemini API on November 19, 2026, one day after opening to Google AI Ultra subscribers — and at roughly $0.10 a second, it is cheap enough for one developer to try.

What Google Genie 4 Actually Is

Google DeepMind announced Genie 4 on October 27, 2026, as the successor to Genie 3. Unlike a video generator, which produces a fixed clip, a world model simulates an environment frame by frame, reacting to what you do: walk forward and the world keeps generating the path ahead; type “make it rain” and the weather changes; ask a character a question and it answers.

The headline feature is sound. Genie 4 is DeepMind’s first real-time world model to generate audio natively, so music, ambient noise, sound effects and spoken dialogue are created at the same time as the visuals, rather than dubbed on afterwards. Characters can move, react and hold conversations that adapt to what you do — the part DeepMind highlights most, because earlier world models were silent and their inhabitants were scenery.

Person immersed in virtual reality wearing VR headset amidst blue and purple lighting
Stepping inside an AI-generated world is the idea behind Genie 4. Stock photo by Michelangelo Buonarroti on Pexels.

Under the surface, Genie 4 leans on the Gemini family. Gemini Omni anchors the simulation in realistic appearance and behaviour, while Gemini 3.8 Flash runs inside the loop as a guiding agent, watching the frames and checking each step. DeepMind showed it off with Gemini 4 Argon, simulating a humanoid robot playing checkers.

Google describes Genie 4 as an early, experimental model. There are no published benchmarks and no independent evaluation so far, and every frame carries an imperceptible SynthID watermark.

Google Genie 4 Gemini API: What Developers Get

The developer release, announced on Google’s blog on November 19, 2026, exposes Genie 4 as models/genie-4 through the v1beta Interactions API, with access also in Google AI Studio. Google calls it the first world model developers can build with.

A session starts with a text prompt, a reference image, or both. From there, Genie 4 streams a continuous world at 1280 by 720 pixels — standard 720p — at 24 frames per second, with synchronised native audio at 48 kHz stereo. You steer with real-time keyboard controls alongside live text prompts, changing the weather, lighting, objects or events on the fly, while Gemini’s reasoning keeps the world state consistent.

Sessions connect over WebRTC for browser apps or WebSockets for server code, starting with one world_setup event, and the API mints single-session ephemeral tokens so the main API key never sits in client-side code.

Google lists three main uses: prototyping interactive experiences, visualising spatial concepts, and training robotics and AI agents in simulated environments. Game developers can sketch a playable level from a sentence; filmmakers can block out a scene before a set is built; robotics teams can rehearse tasks in a world that reacts.

How Much Genie 4 Costs: About $0.10 a Second

The Google Genie 4 Gemini API is not on the free tier. It requires a paid-tier API key, and it bills by the token — except that video output dominates the bill.

Google’s published pricing is $1.50 per million input tokens, $9 per million text output tokens, and $17.50 per million video output tokens, with stereo audio included. At 240 video tokens per frame, a 24-frames-per-second stream uses 5,760 tokens a second — roughly $0.10 per second, or about $6.05 per minute. A two-minute session, the maximum length, costs around $12.

There are limits on how much you can use. Google caps each session at 120 seconds, and its documentation identifies the first 45 to 90 seconds as the strongest period for visual coherence, warning that stability may degrade after 60 to 90 seconds. Sessions are fixed at 16:9, with no portrait mode and no configurable frame rate. Rate limits depend on your paid tier: Tier 1 allows two concurrent sessions and 50 requests a day, Tier 2 allows five sessions and 250 a day, and Tier 3 allows ten sessions and 1,000 a day.

One billing warning: closing a browser tab, or calling session.detach(), does not stop a session — or the billing. Developers must call session.complete() or session.cancel() explicitly, or a $3 experiment can turn into an unpleasant surprise on an invoice.

How to Try Genie 4 Today

There are two ways in, and both are gated. The consumer route is Project Genie, Google’s browser app for creating and exploring generated worlds. Google opened Genie 4 there on November 18, 2026, to eligible Google AI Ultra subscribers, alongside a new Project Genie agent that builds a world from a natural-language description and a community gallery for remixing other people’s worlds. Ultra costs from $99.99 a month. Google AI Pro members, at $19.99 a month, can join the waitlist — and some Pro users have reported receiving access invitations in recent days.

The developer route is the Google Genie 4 Gemini API itself. You need a Google Cloud project with billing enabled and a paid-tier Gemini API key, then the model is available as models/genie-4 in the Interactions API, with starter apps and an interactive canvas in Google AI Studio. Google has not published a date for general availability, and there is no free tier.

You cannot run Genie 4 yourself: there are no weights to download and no local install — it is a closed, hosted model. Anyone offering a Genie 4 download is not legitimate, and should be treated as a security risk.

What Genie 4 Is Good For — and Where It Falls Short

The honest answer is that Genie 4 is brilliant at first ideas and still rough at finished work. For game prototyping, it is genuinely new: a designer can type a description, walk through the result, and iterate in minutes instead of waiting days for a grey-box level. For film pre-visualisation, a director can test a camera move through a location that does not exist. For robotics, a team can rehearse a task in a world that pushes back.

Vivid neon lights create an abstract cityscape, capturing a futuristic urban vibe
A digital illustration of a neon city, not a Genie 4 screenshot. Photo by Pachon in Motion on Pexels.

The limits are equally real. Two-minute sessions are short, and coherence fades after the first minute, so this is not yet a way to generate a finished game or a long scene. There is no portrait mode, which rules out vertical video for phones. No independent benchmarks exist, so claims about realism rest entirely on Google’s own demos.

Remember what Genie 4 is not: it is no game engine. There is no physics to tune, no save file and no deterministic replay — a simulation that improvises, wonderful for brainstorming and unreliable for anything that must behave the same way twice.

Genie 4 vs Genie 3: What Changed

Genie 3, which arrived earlier in 2026, proved a model could generate a navigable 720p world at 24 frames per second in real time. Genie 4 keeps that foundation and adds what its predecessor lacked: sound and steering. Native audio is the biggest change, because a silent world always felt like a demo; mid-session prompting turns the model from a viewer into a collaborator.

Google has also upgraded Genie 3 alongside the launch: generations in Project Genie now run for up to three minutes, up from 60 seconds. The bigger picture is the pace. In the space of a week, Google shipped Gemini 4 Neon for heavy professional work and Gemini 4 Hydrogen for everyday coding, opened Genie 4 to Ultra subscribers, and then put Genie 4 in the Gemini API for developers. The company is releasing its models as a connected set, with Gemini models doing the reasoning and Genie models building the worlds that reasoning plays out in.

The Bottom Line

The Google Genie 4 Gemini API launch matters because it moves world models out of the demo reel and into developers’ hands. For about $0.10 a second, anyone with a paid Gemini API key can now generate a playable, talking world from a sentence — something that, a year ago, existed only in research papers and conference keynotes.

For developers, it is worth a short, careful experiment: start in Google AI Studio, keep sessions brief, and end them properly. For everyone else, the path is a Google AI Ultra subscription or the Project Genie waitlist. Genie 4 is still experimental and unproven outside Google’s own demos — but the price of finding out has never been lower.

Sources: Google, “Genie 4 is now available on the Gemini API”; Google DeepMind, “Genie 4: A new generation of world simulation”.

Leave a Comment