Confirmed 19 Nov 2026 · Labs A one-minute read
Google opened Genie 4 to any developer: 720p interactive worlds from a text prompt.
Genie 4 generates continuous 720p worlds at 24 frames a second, with sound, steered live by keyboard and text. Caveat: it needs a paid key, and the post names no price and no rate limit.
Genie 4 was a research preview for testers and Ultra subscribers.
Any developer with a paid key can build on a world model.
Why it matters
Google lists training robotics and AI agents among the uses. A model that renders a controllable 720p scene is a training ground: an agent can be dropped into a generated environment instead of a real one. SynthID sits in every frame, which tells you Google expects the output to circulate.
Established
- Genie 4 is live in the Gemini API and Google AI Studio.
- Output is continuous 720p at 24 frames a second with native audio.
- Google names robotics and agent training as intended uses.
Not established
- The price. The announcement links to pricing but states no figure.
- Session length or rate limits. Neither appears on Google's post.
- That anyone is training robots on it. No customer is named.
The five-minute readThe record, the timeline, the weight and 3 sources
The record
Google published Genie 4 now available through the Gemini API on blog.google on November 19, 2026, with a published time of 17:00 Universal Time. The post states Genie 4 is available for developers through the Gemini API, by way of the Interactions API, and in Google AI Studio. From a text prompt, a reference image, or both, developers can generate continuous 720p worlds at 24fps with synchronized native audio. Navigation uses real-time keyboard controls alongside live text prompts, and the post says Gemini reasoning keeps the world state consistent and coherent throughout. Every output carries SynthID watermarking, which the post calls imperceptible. The suggested uses listed are prototyping interactive experiences, visualizing spatial designs, and training robotics and AI agents. The post names no price, no rate limit, no session length and no launch partner, and links to pricing documentation instead. The sequence before it: Google DeepMind showed Genie 4 as a research preview to trusted testers in October 2026, naming camera latency and long-term scene permanence as open problems. On November 18 at 18:00 Universal Time Google published the consumer launch, written by Diego Rivas and Pavan Kumar, giving eligible Google AI Ultra subscribers access in Project Genie that day. Project Genie launched in January 2026 with Genie 3 and 60 second generations, for Ultra subscribers in the United States aged 18 and over, and went worldwide in May. Google AI Ultra costs 99.99 dollars a month at its lower tier and 199.99 dollars at its higher one. Google AI Pro, at 19.99 dollars, offers a waitlist only. Reported elsewhere and not on Google's post: a price of about 0.10 dollars per second of streamed output, billed as 240 video output tokens per frame, or 5,760 tokens a second at 24 frames, with stereo audio bundled into the video tokens at no extra charge; a model code of models/genie-4 marked stable; paid tier access only, with no free tier; a 120 second cap on each API session; output at 1280 by 720 in 16:9 landscape, with no portrait and no custom frame rates; and no published concurrency limit.
How it got here
- Oct 27
Google DeepMind shows Genie 4 as a research preview, limited to trusted testers, and names camera latency and long-term scene permanence as open problems.
- Nov 18
Google publishes the consumer launch at 18:00 Universal Time. Eligible Google AI Ultra subscribers get Genie 4 in Project Genie the same day. Pro subscribers get a waitlist.
- Nov 19
The developer post goes up at 17:00 Universal Time: Genie 4 in the Gemini API and Google AI Studio, 720p at 24 frames with native audio, keyboard control, SynthID in every frame.
Game engines, whose asset pipelines this walks around; robotics teams that pay for simulators; Nvidia, which sells Omniverse for the same job; OpenAI and Meta, neither of which has shipped a world model to developers; and anyone who has to tell generated footage from recorded footage.
What to watch
- Would strengthen itA named robotics or autonomy company states that it trains or evaluates policies in Genie 4, or Google publishes a robotics case study using the Gemini API, by March 31, 2027.
- Would strengthen itA second laboratory opens a world model to general developers on paid keys, or Google reports Genie usage figures, by June 30, 2027.
- Would weaken itGenie 4 is withdrawn from the Gemini API, restricted back to approved partners, or still has no published rate limits and no named customer by June 30, 2027.
Why it weighs this much
Cost is medium. Serving a continuous video stream at 24 frames a second is expensive to run, and the post commits Google to doing it for strangers rather than for testers. Commitment is medium, not high: an API in the Gemini surface is a product decision Google makes often, and the post carries no customer, no price and no service promise. Reversibility is easy, which is the honest score. Google has withdrawn preview models before, and nothing here is contractual. Persistence is a year: a model code marked stable and a documented interface tend to outlive the news cycle, but world models are moving fast enough that Genie 4 is unlikely to be the current one in 12 months.
The words
- World model
- A model that generates a scene you can move through, frame by frame, rather than a single image or a fixed clip.
- Interactions API
- Google's interface for sessions that run continuously, where the model holds state while the user acts on it.
- SynthID
- Google's invisible watermark, written into generated output so it can later be identified as machine made.
Sources
- BGoogle: Genie 4 now available through the Gemini API, blog.google, Nov 19, 2026, published 17:00 Universal Timeblog.google
- PPPC Land: Google charges developers about 10 cents a second for Genie 4 worlds, Nov 19, 2026ppc.land
- DDatastudios: Project Genie, Ultra access and the Nov 18 consumer launch, checked Nov 19, 2026datastudios.org
Your call
Which explanation do you believe?
A call is final. Other people's calls show after yours.