blog.google web signal

Google DeepMind Ships Genie 4 With Mid-Session Scene Edits

5 sources tracking this story

TL;DR

  • Genie 4 generates native audio including ambient sound, dialogue, and music integrated directly inside the world model, not layered on after rendering.
  • Gemini 3.8 Flash runs as a live planning agent inside Genie 4, breaking user goals into sub-steps and adapting plans as the scene changes.
  • Gaming stocks crashed 8-24% on Genie 3's February 2026 launch then recovered within days, giving investors a calibration point for Genie 4 market reaction.

Google DeepMind let subscribers edit a generated world while it runs. In a launch post on blog.google, the team describes Genie 4 as "our most realistic and interactive world model yet," one that lets a user "prompt mid-session to change the world you created."

The pitch, in the post's own words: "Shift the weather, drop in a character, or redirect the action." Genie 4 also pairs the visuals with "real-time native audio that adapts as the scene unfolds," from ambient soundscapes to character dialogue.

Alongside Genie 4, Google is pushing updates into Genie 3. Sessions now run "up to 3 minutes per generation (up from 60 seconds)," and users can link worlds together for longer journeys or seed a scene from a 360-degree panorama image. A new Project Genie agent conducts the build conversationally. "Now crafting an interactive universe is as effortless as having a conversation," the post reads, and the agent can call Gemini models to surface context-aware suggestions mid-scene.

Access is tiered. Eligible Google AI Ultra subscribers get Project Genie starting today; AI Pro members can join a waitlist. Published creations land in a community gallery, alongside collaborations the team highlights with BLVCKL!GHT ("Villa Giallo") and filmmaker Anand Gandhi ("The Life of Ants"). All generated video and audio carry SynthID watermarking, and the post lists known limitations around camera latency, complex physical interactions, audio precision, and scene permanence.

Genie 4 lands into a dense stretch for generative video on our video-generation tracker, which logged 55 stories across the prior 90 days.

What others are reporting

Coverage cluster as of 8h after publish

  1. Google DeepMind Read →

    First-party post reveals Gemini 3.8 Flash as a live planning agent inside Genie 4, confirms native multimodal audio generation, and notes SynthID watermarks on all outputs.

    You can set an event in motion, and Genie 4 simulates what happens next: how the world responds, how it sounds, and how the people in it move, react, and speak.
  2. TechCrunch Read →

    Places Genie 4 inside a consolidating market, citing AMD's World Labs acquisition, Amazon's Decart talks, and Rivian's Third Dimension purchase. Notes object drift and text legibility as unresolved limits.

    What we'll increasingly see are hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time.
  3. Startup Fortune Read →

    Documents Genie 3's February 2026 crash: Unity fell 22-24%, Roblox 12.8%, Take-Two 8-10%, all recovering within days, providing a historical template for how investors should read Genie 4.

    Project Genie's output was, in his words, unsuitable on its own for games requiring consistent, repeatable player experiences.
  4. DEV Community Read →

    Details US-only access at the $250/month AI Ultra tier, highlights new multiplayer mode, and positions Genie 4 as a prototyping tool rather than a production game-engine replacement.