Google launches Genie 4 world model with audio and two-minute interactive sessions, powered by Gemini 3.8 Flash
Google announced Genie 4, a world model for generating interactive AI experiences that run about two minutes, up from one minute for Genie 3. It adds audio generation, real-time natural language updates and Gemini 3.8 Flash for its simulations. It launches as a research preview for trusted testers.
Google announced Genie 4 on October 27, 2026, a world model that generates interactive experiences and can now run them for about two minutes, up from one minute for the previous Genie 3 model. It is launching first as a "research preview" for "trusted testers," according to The Verge's Jay Peters.
What changed from Genie 3
Google says Genie 4 has four main updates:
- More realistic worlds. Google says the model produces more realism than its predecessor.
- Real-time prompt updates. Users can change an experience while it runs using a natural language prompt.
- Audio generation. The model can create audio to accompany its interactive worlds.
- Gemini 3.8 Flash. Genie 4 uses Gemini 3.8 Flash under the hood to help with its simulations.
Session length is about two minutes, compared with one minute for Genie 3.
Intended uses and stated limits
Google is pitching Genie 4 for pre-visualization for filmmakers, prototyping game ideas and building training scenarios for robots. The company also warns that as simulations run longer, details and object permanence can shift over time.
Availability
Genie 4 is limited to trusted testers in a research preview. Google says it will be "coming soon" to Google AI Ultra subscribers, and it has opened a waitlist for people who want to try it in Project Genie. Peters notes that Project Genie was powered by Genie 3. Peters had not been able to try Genie 4 himself, so he could not compare it with the Genie 3-powered version. The examples in his report come from Google's own demonstrations.
Parameter count, context window, benchmark scores and pricing have not yet been disclosed.
What this means
The two-minute session length and the admitted drift in details and object permanence show that long-horizon consistency is still the main constraint on this class of model. Real-time prompt editing and audio move Genie 4 from a generated visual environment toward something closer to an interactive scene that can be steered while it plays. Until the model reaches Google AI Ultra subscribers or a wider audience, the realism and consistency claims rest on Google's own demos, and the stated use cases for film, games and robot training remain intended applications rather than demonstrated results.
Related Articles
Gemini 4 Argon's 11-hour unattended run yields two accepted optimizations in Google's ADK codebase
Google says an 11-hour unattended run of Gemini 4 Argon produced two accepted optimizations for its Agent Development Kit (ADK), cutting in-memory session-preparation overhead by about 40% in long sessions and cold package import time by 21%. Argon also produced 4 modifications that developers rejected on review.
Google's Data Agent Kit reaches GA with 37 open-source skills and an IAM deny policy for read-only agents
Google has made Data Agent Kit generally available. It is free, open source, and installs into Antigravity, Claude Code, Codex, Cursor, and VS Code-compatible editors. It splits guardrails between advisory skills and enforceable IAM deny policies, according to InfoQ.
Gemini Notebook becomes a Connected App, letting users manage notebooks from inside Gemini chats
Google has added Gemini Notebook as a Connected App in the Gemini app, so users can manage and query notebooks directly within chats using natural language. The integration is on by default globally, except in the European Economic Area, where it is off by default.
Mirendil releases Fern 0.1, its first model, with a 10M-token context window
Mirendil has released Fern 0.1, its first model, pretrained from scratch with a 10M-token context window. It is aimed at agentic AI R&D work and scores 100% on RULER single needle at 10M, but trails frontier models on agentic coding, according to MarkTechPost.
Comments
Loading...