Closed SourceGoogleReleased 2026-11

Genie 4

Google DeepMind's real-time interactive world model, now on the paid Gemini API

Context window

Not published (sessions up to 120 seconds)

Input / 1M tokens

$1.50

Output / 1M tokens

$17.50

Provider

Google

Paid tier only, per 1M tokens: $1.50 input (text or reference image), $9.00 text output and $17.50 video output with stereo audio included. Video is billed at 240 output tokens per frame (5,760 tokens per second at 24 fps), about $0.10 per second or $6.05 per minute. No context caching, free tier, Batch, Flex or Priority. Billing runs while a session is active, until complete() is called. · Data verified 2026-11-20

Genie 4 is Google DeepMind's world model. From a text prompt or a 16:9 reference image it generates an interactive world as 720p, 24 fps video with synchronized 48 kHz stereo audio, which the user moves through with keyboard and camera controls and can change mid-session with new text prompts, such as different weather or a new character. Google calls it its most realistic and interactive world model yet and the first in the Genie line with real-time native audio. On November 18, 2026 it opened to eligible Google AI Ultra subscribers in Project Genie, and the Gemini API changelog lists it as generally available on the paid tier from November 19 (model ID genie-4) through the Interactions API over WebRTC or WebSockets. Sessions last up to 120 seconds, with the strongest visual coherence in the first 45 to 90 seconds, and video output costs about $0.10 per second.

How to access it

Gemini API paid tier only (model ID genie-4), generally available since November 19, 2026 through the Interactions API with WebRTC or WebSocket streaming and ephemeral session tokens; served from the United States, Western Europe and Japan. Eligible Google AI Ultra subscribers can use it in Project Genie (labs.google/projectgenie), where Google AI Pro members can join a waitlist.

Strengths

  • ✓Real-time interactive worlds with synchronized stereo audio, including ambient sound and character dialogue
  • ✓Text steering during a session: change the weather, add a character or redirect the action
  • ✓Starts from a text prompt or a single 16:9 reference image
  • ✓SynthID watermarking on all generated video and audio

Best for developers who...

Prototyping game worlds and levelsInteractive demos, stories and educational scenesApps that stream steerable video to users in real time

When to choose it (and when not to)

Reach for Genie 4 when...

  • →You need a playable, steerable world rather than a fixed video clip
  • →You want generated video and audio in one real-time stream

Look elsewhere if...

  • ✕Experiences longer than 120 seconds per session
  • ✕Fixed clips for editing, ads or social video (Gemini Omni Flash is built for that)
  • ✕Free-tier or batch use: Standard paid tier only, with no Batch, Flex or Priority
  • ✕Precise physics or long-term scene permanence, which Google lists as open challenges

How to use it

  • ›Call complete() to end a session; billing continues while a session is active
  • ›Send one 16:9 reference image, kept under 48 KB as a compressed JPEG on WebRTC
  • ›Put the important moments in the first 45 to 90 seconds, where Google says visual coherence is strongest

Quickstart

Note
Open a Genie 4 session with the Interactions API (model genie-4): from Python with google-genai over WebSocket, or from JavaScript with @google/genai over WebRTC or WebSocket. Send the world setup (a text prompt and optional 16:9 reference image), stream user_action controls and text steering, then call complete() to end the session.

See Google's Genie 4 model page for session setup and ephemeral tokens.

API model id: genie-4

Compare Genie 4

Compare Genie 4 with any other model

Build a comparison →
All model comparisons →

Learn the concepts