Google opens Genie 4 to developers at about 10 cents a second
The world model streams 720p at 24 fps with synchronized stereo audio over the Interactions API, and a paid-tier Gemini API key is required because there is no free tier.

Key takeaways
- Google opened Genie 4, its real-time interactive world model, to developers on November 19, 2026, through the Gemini API's Interactions API and Google AI Studio.
- A Genie 4 session needs a paid-tier Gemini API key and bills about $0.10 per second of streamed video, at 240 video output tokens per frame and $17.50 per 1 million video output tokens with stereo audio included.
- Sessions run for up to 120 seconds, stream 720p video at 24 fps over WebRTC or WebSocket, and are seeded from a text prompt, one 16:9 reference image, or both.
- Google's Tier 1 rate limits allow 2 concurrent Genie 4 sessions and 50 requests per day, and a client must call complete() to stop billing because closing a browser tab does not stop server generation.
Google opened Genie 4, its real-time interactive world model, to developers on November 19, 2026, through the Gemini API's Interactions API and Google AI Studio. Any developer with a paid-tier Gemini API key can now run the model as genie-4 and stream a steerable 720p world with synchronized stereo audio.
SLOP TV covered the Genie 4 launch on October 27, when Google DeepMind introduced it as a research preview for trusted testers and said developer access would follow. What changed here is access and price: Google's developer post confirms general availability on the Interactions API, and the model's own documentation now carries the specs, session cap and rate limits a builder needs.
What Genie 4 does that a clip generator does not
A world model generates a playable environment rather than a fixed clip, so the scene keeps running and responding while you move inside it. Google's Genie 4 guide contrasts it with offline generation: whereas Gemini Omni 1.2 Flash renders a finished 3-to-15 second clip, Genie 4 renders frames on the fly in response to live keyboard movement and to natural-language events sent mid-stream.
That frame-by-frame design has a consequence for anyone hoping to use it as a render farm. Genie 4 simulates over a rolling temporal window rather than keeping a persistent 3D scene graph, so the world is not a place you can save and return to. Looking away from an object for eight seconds or more, or turning a full 360 degrees, can make landmarks, doors or buttons shift or reset, and the model page puts the strongest visual coherence in the first 45 to 90 seconds.
How to start a Genie 4 session
The route is a paid-tier Gemini API key, the Interactions API, and Google's GenAI SDK. The setup guide walks through a browser client that connects over WebRTC with @google/genai, or a Python or Node.js backend that connects over WebSocket with google-genai.
A session begins with one world_setup event carrying a world_prompt text string, a single 16:9 starting image in a reference_image field, or both. The docs advise keeping that seed image under 48 KB as a compressed JPEG over WebRTC, because the event has to fit inside the 64 KB limit on the SCTP data channel. Google says initial world setup takes roughly 5 to 9 seconds before the first frame arrives.
In a browser, WebSockets need an ephemeral auth_tokens value rather than the raw API key, because standard browser WebSocket APIs cannot attach the x-goog-api-key header. A backend mints that token, which keeps the main key off the client.
How to steer a world while it runs
Once connected, the client drives the simulation with JSON events. Native WASD and arrow-key user_action events move and pan the camera and land in about 0.2 to 0.5 seconds. Mid-session text input events, which change weather, lighting, objects or events, pass through world planning and take about 1 to 2 seconds to appear on screen.
Two keys in the wire schema, key_e and space, exist but Genie 4 attaches no built-in jump or interact behavior to them, according to the docs. Custom controls are built by catching those keys in the client and sending descriptive text events instead. Google's prompt and controls guide lists five tested patterns for that, including cause-and-effect rules written as Rule clauses inside the world prompt.
What a Genie 4 session costs
Genie 4 is not on the free tier. Google's pricing page lists $1.50 per 1 million input tokens for text or a reference image, $9.00 per 1 million text output tokens, and $17.50 per 1 million video output tokens with stereo audio included. Billing runs on emitted video at 240 tokens per frame, which is 5,760 tokens per second at 24 fps, an effective rate of about $0.10 per second of streamed output, or roughly $6.05 per minute.
That meter keeps running until the session is closed. The docs warn that closing a browser tab detaches the transport without stopping server generation, so a client should always call complete() to end the session and stop billing.
The limits Google documents
A Genie 4 session caps at 120 seconds. Output is fixed at 1280x720, 16:9 and 24 fps; portrait 9:16 output and configurable frame rates are not supported. The model page also lists no live microphone or webcam input, no function calling, no Google Search grounding, no structured outputs, and no session resumption across completed interactions.
Rate limits scale with the project's usage tier. Google's rate-limits page gives Tier 1 two concurrent sessions and 50 requests per day, Tier 2 five and 250, and Tier 3 ten and 1,000. Control messages sent inside a session do not consume that daily quota. Sessions are served from data centers in the United States, Western Europe and Japan, with round-trip latency under 150 milliseconds in North America, Western and Central Europe, and East Asia.
Safety filtering runs during the session too. The docs say prompts or reference images containing identifiable real people, public figures or copyrighted brand IP can be blocked at startup or terminate an active stream mid-session. Every output carries an imperceptible SynthID watermark, Google's developer post says.
What it means for video makers
Genie 4 is not a replacement for a clip generator, and the specs are the reason. A 120-second cap, a rolling horizon and text that deforms over time make it a prototyping tool rather than a shot source. The practical fit for a video maker is previsualization and interactive spatial design, two of the uses Google names alongside robotics and agent training. A director can walk a set before it exists; an editor cannot render a finished 10-second beat and cut it in.
The price settles the comparison. At about $0.10 per second, a full session runs to roughly $12, and a failed experiment costs the length of the session rather than a render queue. Whether that is worth it depends on how a team uses the output, because Genie 4 does not hand back a file.
Get a paid-tier key in Google AI Studio, open the Live playground with the Genie 4 model to test a world prompt and its action bindings, then follow the setup guide to connect a client over WebRTC or WebSocket.
Sources
- blog.google - Google's November 19 developer announcement and SynthID note
- aistudio.google.com - Genie 4 setup, transports, event protocol, session and safety limits
- aistudio.google.com - Genie 4 model specs, stream limits and unsupported features
- aistudio.google.com - Genie 4 token pricing and the effective per-second rate
- ai.google.dev - Genie 4 concurrent session and daily request limits by usage tier
- aistudio.google.com - Genie 4 prompt and control patterns
- ai.google.dev - Interactions API availability and surface
- ppc.land - independent summary of the developer release and the Project Genie split
- testingcatalog.com - corroborates the price, the 120-second cap and the launch day
- sloptvnews.com - SLOP TV's October 27 launch coverage, the follow-up baseline