Access is the news

When DeepMind announced Genie 4 on October 27, the model existed for trusted testers. That changed today. Google is rolling the new Project Genie experience out to eligible Ultra subscribers, which makes Genie 4 a model some paying readers can actually try rather than only watch in a demo reel.

Two words in Google's announcement deserve attention. The first is "eligible." Google does not publish which Ultra accounts receive the new experience today, in which countries, or on what schedule. The second is the tier itself. Pro members get a waitlist. Genie 4 remains a top-tier subscription feature, and the announcement gives no usage limits for it, so what an Ultra subscription actually buys in world generation is unspecified.

This piece is separate from our October 27 coverage of the Genie 4 announcement, which dealt with a research preview and the specifications DeepMind did not publish. This one deals with a product opening its doors, and with what the opening reveals.

The three-minute number belongs to the older model

Read the announcement closely and its only hard performance figure sits in the Genie 3 section. Sessions for Genie 3 now run up to three minutes per generation, triple the previous 60 seconds, and users can chain worlds together for longer journeys or seed one from a 360-degree panorama image.

Genie 4 gets no equivalent figure. Google describes Genie 4 as its highest-fidelity option, with better visuals, physics and native audio, claims the post supports with no measurements, and then directs readers who want "longer runs" to Genie 3 in the model picker. Two different limits are in play, and the post keeps them separate even if its readers may not. Session length is a number Google published for Genie 3 and not for Genie 4. Scene permanence, how long a world holds together as itself, is a challenge Google names separately, in its closing list of unsolved problems: camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence. Until Google publishes a Genie 4 duration or a permanence result, what the flagship's worlds can sustain is unknown, not merely short.

The rest of the update is about turning a research model into a place people return to. The Project Genie agent builds through conversation: it orchestrates Google's multimodal models to assemble initial visuals and a custom background score from a description. Separately, Gemini models can join a session as the user explores, observe the screen, surface suggestion chips, or trigger in-world events automatically. The gallery adds the social layer, with curated worlds, remixes and user publishing, including collaborations with a creative technologist and a filmmaker.

That structure will be familiar to anyone who watched image generators make the same trip, from model page to canvas to community feed. The open question is whether a gallery can hold attention when the worlds in it run on models whose long-term scene permanence their maker lists as unsolved. Sharing a world is a different proposition from sharing a picture when the world itself is the part still under research.

Sources

  1. [1] Google blog, "Project Genie update featuring Genie 4," November 18, 2026 — the primary source, read first-hand in full: the Ultra rollout and Pro waitlist, Genie 4 mid-session prompting and native audio, the Genie 3 three-minute sessions, world chaining and panorama grounding, the agent and gallery, SynthID watermarking, and the named unsolved problems.Read source
  2. [2] Unite.AI, "Google Brings Genie 4 to Project Genie for Eligible AI Ultra Users," November 18, 2026 — timing confirmation only.Read source
  3. [3] alfw.ai, "Genie 4 Adds Sound and Speech to Google's Worlds. DeepMind Published No Specs," October 27, 2026 — earlier announcement coverage, for the distinction above.Read source