Google showcases a 3D world built in minutes

Google AI’s post, timestamped October 15, invited viewers to see a fully explorable 3D world build itself in minutes.[1] The showcased scene was a multi-tiered pagoda sanctuary, and the post invited viewers to watch Google DeepMind’s GeokenAI explain the demonstration.[1] The post’s invitation is a demonstration teaser; the supplied material does not give a build time beyond “in minutes” or describe the scene’s technical specifications. For teams evaluating generated 3D work, the scene shows the kind of output being presented, but does not establish how it performs across other projects.

Google’s AI Studio publication, dated October 13, says engineers and creative technologists across Google pushed Gemini 4 Argon across hundreds of 3D builds.[2] Google says Argon has a 1M-token context and output window and can plan, write large multi-file 3D codebases and iterate autonomously.[2] It can interpret photos, sketches, concept mockups and screenshots of its rendered scenes, according to Google.[2] Google also says the model handles coordinate math, physics, camera optics and mesh geometry under the hood.[2] These capabilities point to a workflow where a team can provide visual references and refine a project through iterations.

Google lists browser-based uses including self-contained 3D games, procedural scenes, interactive WebGPU shaders and fast prototypes.[2] It also describes multi-file games, long-running goal loops, parallel subagents, browser screenshot checks and local Blender pipelines.[2] Google advises users to orchestrate parallel subagents and says its Critic subagent can provide critical feedback.[2] In a Desk Racer example, Argon placed 68 coaster pillars without clipping into the desk or props; the creator also asked it to shift to 1:150 scale and move the chase camera closer.[2] Access is for developers with Google AI Ultra plans or paid Gemini API keys, and Ultra subscribers can select Argon in Antigravity.[2]

Google DeepMind separately announced Genie 4 as a world model intended to improve real-time interactive simulation.[3] It says Genie 4 can generate music, ambient noise and interactive dialogue alongside a simulated world, and can be steered in real time with natural language to change weather, move the camera or introduce a person or object.[3] Google DeepMind said Genie 4 had started rolling out as a research preview for trusted testers.[3] Bloomberg reported that Genie 4 lets users create explorable worlds from text prompts and navigate them with keyboard controls.[4] A commenter questioned whether anyone wanted to use generated worlds, while Riley Goodside called Genie 4’s progress over Genie 3 striking.[1][5]