What changed

Google released Genie 4 (models/genie-4) on November 19, 2026, according to its Gemini API release notes. Developers can seed interactive worlds from text, a 16:9 reference image, or both, and steer them live using WASD keys, arrow keys, or mid-session text events. The model streams over WebRTC and WebSockets, the release notes say.1

How it works

According to Google's Genie 4 documentation, each session follows a four-step lifecycle on the v1beta Interactions API. A backend server creates the interaction and mints a single-session ephemeral auth token so the main API key never reaches the browser. The client then connects over WebRTC for browser apps or WebSockets for Python and Node.js pipelines, sends one world_setup event, and drives movement with user_action or input events, the guide explains.23

For a real browser session, builders must call complete() when a user finishes, because closing a tab or detaching the transport does not immediately stop server-side generation or token billing, Google's guide warns. Output streams at 1280x720 resolution and bills at 240 video output tokens per frame, or 5,760 tokens per second, with stereo audio bundled in. On WebRTC, reference images must stay under 48 KB as compressed JPEGs to fit inside the 64 KB SCTP data channel limit, the model page states.23

Responsiveness depends on both control type and geography, according to the interactive world generation guide. Initial world setup takes roughly five to nine seconds before the first frame arrives, while native WASD and arrow keys respond in about 0.2 to 0.5 seconds and mid-session text prompts take one to two seconds. Google serves Genie 4 from data centers in the United States, Western Europe, and Japan, where round-trip latency stays under 150 ms; connections from other regions face slower steering and potentially lower frame rates.3

Paid-tier quotas scale up to Tier 2's five concurrent sessions and 250 daily requests, and Tier 3's 10 concurrent sessions and 1,000 daily requests, according to Google's rate limits page, while in-session control events do not consume quota. Google's documentation notes several limits: visual coherence is strongest in the first 45 to 90 seconds because the model uses a rolling temporal window rather than a persistent 3D scene graph, so turning away can cause objects to shift. Genie 4 does not support live microphone or webcam input, function calling, session resumption, or Batch API jobs, and Google's pages do not explain when text output tokens are billed alongside video output.234