Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Genie 4 in Project Genie: Who Can Try It and What Changed

Google opened Genie 4 in Project Genie on November 18, 2026: edit worlds mid-session with native audio. Ultra only, Pro waitlist; Genie 3 runs longer.

RA
· Founder, AIToolRanked
Frame from Google's Genie 4 demo The Life of Ants: black ants carrying crumbs around two tunnel entrances in a mossy anthill, with dewy grass and a dry leaf behind
On this page

What Google rolled out on November 18

On November 18, 2026, Google published Genie 4 is now available ↗, by Diego Rivas (Google DeepMind) and Pavan Kumar (Google Labs). It rolls out a new Project Genie experience to eligible Google AI Ultra subscribers, with first-time access to Genie 4, upgrades to Genie 3, a Project Genie agent, and a community gallery. Google AI Pro members can join a waitlist on the same page. Google DeepMind's Genie 4 model card ↗ carries the same date. The Next Web ↗ covered the rollout the same day. The lead image is a frame from Google's "The Life of Ants" demo video in the launch post, used for news reporting; credit Image: Google. We have not used Project Genie.

Project Genie is a browser prototype where you generate a world and then move through it. Genie 4 adds two things Google leads with. You can type a new prompt mid-session to change the world: shift the weather, drop in a character, or redirect the action. And for the first time, the visuals come with real-time native audio, "from ambient soundscapes to character dialogue."

Google first announced Genie 4 ↗ on October 27 for trusted testers. That post said it was coming to Ultra subscribers in Project Genie and to "early access partners in the Gemini API and AI Studio." The November 18 rollout is the Project Genie part. It does not open an API.

What is in the model card

The model card describes Genie 4 as built on Gemini Omni Flash, with Gemini handling reasoning and real-time steering. Inputs are text, images, and live keyboard and mouse actions. Output is "Real-time 720p video at up to 24 frames per second, with synchronized audio." Google lists the limits itself: visual details and object permanence can drift as a session runs longer or after several complex prompt edits. Complex physics, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency are listed as open research problems.

Genie 3 stays in the model picker, and Google says to pick it "when you want longer runs or panorama-seeded spaces." The update gives Genie 3 runs of up to 3 minutes per generation (up from 60 seconds), world chaining, 360-degree panorama grounding, and real-time in-world events. Google gives no session length for Genie 4.

Genie 4Genie 3
Native audioYesNot listed
Prompt mid-sessionYesYes, as events
Run lengthNot statedUp to 3 min
Panorama seedNot listedYes

The new agent builds a world from a conversation. Google says it "orchestrates Google's suite of multimodal models" for the world's look and a background score. Gemini can join a session, watch the screen, offer suggestion chips, or trigger events. The gallery holds curated and community worlds you can explore and remix, including "Villa Giallo," made with creative technologist Will aka BLVCKL!GHT, and "The Life of Ants," made with filmmaker Anand Gandhi. Google says all video and audio made in Project Genie carries SynthID watermarking, and inputs and outputs go through safety screening.

Who can actually try it

Google's blog says "eligible Google AI Ultra subscribers". Google's AI plans page ↗ is more specific. It says Project Genie is available in more than 140 countries, and its footnote reads: "Only available with a Google AI Ultra 20x plan." The same page lists an Ultra 5x tier below it. If you pay for Ultra and do not see Project Genie, check which Ultra tier you are on. Pro members get the waitlist, not access.

What people are saying

  • "Today we are launching Genie 4 in Project Genie, alongside some major new updates across the creation experience and gallery!" — Diego Rivas ↗ (@drivascos), Group Product Manager at Google DeepMind and co-author of the launch post.

  • "most people think ai can't make gta yet, but google is getting closer every update" — Kawai ↗ (@kawaidgaf)

  • "Honestly this one feels like sci-fi." — ZarpeMan ↗ (@ZarpeMan)

  • "But the longest runs in today's update aren't Genie 4's." — BroadwakeAI ↗ (@BroadwakeAI)

Reddit was not reachable for this story.

Our take

It is a showcase, not a game engine. 720p at up to 24 frames per second, no published session length for Genie 4, and Google's own warning about drift and object permanence put this in the prototype bucket. The GTA comparison on X is enthusiasm. TNW notes that big studios remain wary, and that Take-Two says GTA 6 used no generative AI.

Mid-session prompts and audio are the real step. Changing the weather or adding a character while you walk through a world, with sound that follows, is what separates this from a generated clip. BroadwakeAI's catch is fair: if you want the longest runs, Google itself points you to Genie 3.

Who should care. Ultra 20x subscribers who already pay for the tier, creators who want to prototype a scene or a mood, and anyone tracking world models. Everyone else is on a waitlist or outside the plan. Developers should wait for the API partners Google mentioned in October. For alternatives, World Labs Atlas is open to everyone with a developer API, and NVIDIA's Cosmos 3 ships open weights for simulation work. For short generated clips rather than worlds, see Gemini Omni 1.2 Flash and our AI video generators list. This week's Google model is Gemini 4 Neon.

Verdict: Try it if you already have Ultra 20x. Do not upgrade for it alone. Watch for Genie 4 session lengths, the Pro waitlist, and API access.

More from the news desk · About Rai Ansar