Google releases Genie 4 world generation through the Gemini API
Genie 4 generates steerable worlds from text or images at 720p and 24fps with native audio. Developers can stream interactive sessions through WebRTC or WebSocket using Google's Live Interactions API.

TL;DR
- Genie 4 is available through the Gemini API and Google AI Studio, following the developer announcement.
- Text or images seed steerable 720p worlds at 24fps with 48 kHz stereo audio, according to the launch specifications.
- Navigation and live prompts can change weather, introduce characters, and reshape a running scene, as demonstrated in the world-editing demo.
- Live Interactions supports WebRTC with ephemeral browser tokens and WebSocket, with Genie 4 its first model, according to the API announcement.
- Video output costs about $6.05 per minute under Google’s pricing; the API walkthrough explicitly calls
complete()to stop generation and billing.
The API guide repeats short action prompts ten times to catch a planning window, while Space has no built-in jump behavior despite appearing in the schema. Google’s model card identifies Gemini Omni Flash underneath Genie 4, with Gemini supplying reasoning and real-time steering.
What shipped
- Gemini API:
models/genie-4on thev1betaInteractions API, released November 19, according to Google’s launch post. - Google AI Studio: an interactive canvas and starter apps; access requires linking a paid-tier API key, as shown in the Studio screenshot.
- Starting conditions: a text prompt, a reference image, or both.
- Output: fixed 1280×720, 16:9 video at 24fps, with synchronized native audio covering ambience, effects, and character dialogue.
- Watermarking: every output includes imperceptible SynthID watermarking, according to the launch post.
Benchmarks and evaluation
Google describes human and automated evaluations in its model card, covering four dimensions:
- Visual quality and aesthetics.
- Environmental realism.
- Audio quality and synchronization.
- Responsiveness to user controls and text steering.
The card reports improvements over Genie 3 in photorealism, fluid movement, and steering. It publishes no numerical scores, sample counts, or reproducible prompt sets; neither the card nor the API launch post reports third-party benchmark results.
The published video specifications remain unchanged between Genie 3’s original release and Genie 4’s release specifications:
- Vertical resolution: 720 pixels → 720 pixels, +0%.
- Frame rate: 24 fps → 24 fps, +0%.
Google lists robotics and AI-agent training among potential applications in the API launch post, without reporting downstream training results.
Regressions and limitations
Genie 3’s consumer session limit increased from 60 seconds → 180 seconds, +200%, in the November 18 Project Genie update. Genie 4’s API guide caps sessions at 120 seconds, 60 seconds shorter than the current Genie 3 consumer limit.
Google’s API announcement says Gemini reasoning keeps the world state coherent throughout interactions. The guide describes a rolling visual horizon with narrower persistence:
- Landmarks, doors, or buttons can shift after being out of view for more than eight seconds or after a full rotation.
- Visual stability degrades after roughly 60–90 seconds.
- Turning away and returning can rearrange buildings; the system does not maintain a persistent 3D mesh.
- Live microphone and webcam input are unsupported.
- Tool calling, Google Search grounding, and structured outputs are unsupported.
- Safety filters can block identifiable people or copyrighted brand IP at startup or terminate an active stream.
WebRTC and WebSocket
WebRTC and WebSocket carry different media formats in the API guide:
| Transport | Media delivery | Transport behavior |
| --- | --- | --- |
| WebRTC | Native RTP video and Opus audio tracks, rendered through a browser <video> element | UDP media with browser jitter buffering; JSON controls use an SCTP data channel |
| WebSocket | JSON step.delta events carrying base64 JPEG frames and 48 kHz, signed 16-bit little-endian stereo PCM | About 33% base64 expansion; TCP head-of-line blocking under packet loss |
WebRTC control messages have a 64 KB limit. The guide specifies compressed JPEG seed images below 48 KB before base64 encoding; WebSocket accepts larger payloads.
Serving locations are the United States, Western Europe, and Japan. Google warns that clients farther away can experience slower steering and reduced frame rates.
Session lifecycle
A session follows four steps in Google’s quickstart:
- Create: the backend calls
interactions.createwithmodels/genie-4. For browser access, it mints an interaction-scoped token throughinteractions.authTokens.create. - Connect: the client calls
interactions.connectusing WebRTC or WebSocket. The browser receives the ephemeral token rather than the long-lived API key. - Initialize: send exactly one
world_setupevent, followed byuser_actionmovement events orinputtext events. A seed image goes in the top-levelreference_imagefield, alongsideworld_prompt. - Finish:
await session.complete()terminates generation and billing.session.cancel()immediately aborts generation.
Closing a browser tab, calling detach(), or exiting Python’s async with disconnects the transport without explicitly completing the server interaction. Genie 4 does not support session resumption.
Browser WebSockets also require the ephemeral token: passing a raw API key to that browser transport throws a usage error by design.
Live controls
Movement and world editing use separate mechanisms in the control protocol:
- Movement: WASD maps to forward, backward, and strafing actions. A held key continues applying until its
key_uparrives. - Camera: arrow keys control looking up, down, left, and right.
- Custom actions: schema values such as
key_eandspacehave no built-in interact or jump behavior. The sample client translates custom keys into descriptive text prompts. - Text edits: planning samples inputs at four-frame step boundaries. Google’s sample repeats a short prompt ten times at 150 ms intervals, with roughly a two-second cooldown before a conflicting prompt.
- Dialogue: quoted speech in a text input can trigger spoken character dialogue when facing a character.
Google documents different response times for each stage:
| Stage | Documented latency |
| --- | --- |
| Initial world setup | Approximately 5–9 seconds before the first frame |
| Native movement | Approximately 0.2–0.5 seconds at low network RTT |
| Mid-session text changes | Approximately 1–2 seconds |
Pricing and quotas
Google’s pricing table charges separately for inputs, text output, and generated video:
| Category | Price per million tokens |
| --- | --- |
| Text and reference-image input | $1.50 |
| Text output | $9.00 |
| Video output, including stereo audio | $17.50 |
Video consumes 240 tokens per frame, or 5,760 tokens per second at 24fps. That works out to $6.048 per minute, or about $12.10 for 120 seconds of output, before input and text-output charges.
The rate limits apply per project:
- Tier 1: two concurrent active sessions and 50 requests per day.
- Tier 2: five concurrent sessions and 250 requests per day.
- Tier 3: ten concurrent sessions and 1,000 requests per day.
Movement and mid-session text events do not consume separate request quota. Genie 4 has no free tier, Batch API, Flex inference, or Priority inference.
Project Genie access
Consumer access began with eligible Google AI Ultra subscribers, as described in the earlier rollout report. Google’s consumer announcement put Pro members on a waitlist.
Selected Pro users subsequently received access invitations, according to testingcatalog’s report. The invitation grants access through the Pro subscription, separately from paid Gemini API usage.