Key takeaways
- Google announced Genie 4 developer availability on November 19, 2026 through the Gemini Interactions API and Google AI Studio; the model ID is genie-4.
- The paid-tier model generates live 1280×720 video at 24 fps with synchronized 48 kHz stereo audio, controlled by keyboard inputs and mid-session text.
- Google lists $17.50 per million video output tokens including audio, approximately $6.05 per minute at the documented frame rate, plus applicable input and text-output charges.
- Sessions have a 120-second cap, no resumption and limited object permanence. Explicitly complete or cancel a session: disconnecting alone does not reliably stop billing.
What Google made available
Google announced on November 19, 2026 that Genie 4 is available to developers through the Gemini API’s Interactions API and Google AI Studio. The documented model identifier is genie-4. A text prompt, one reference image or both seed a continuous world that responds to movement and camera controls, while live text prompts can change the scene. This is paid-tier API access, not a free model or a downloadable-weights release.
The change is developer access, rather than the first unveiling of the research model. Our earlier Genie 4 briefing covered the research preview; Google’s November 18 Project Genie announcement separately describes a consumer rollout for eligible Google AI Ultra subscribers. API builders should use the paid-key requirements and rates in the developer documentation, not assume a consumer subscription includes API usage.
Google advertises interactive experiences, spatial design visualization and experimentation with robotics or AI agents. Those are proposed applications, not evidence that generated environments provide accurate physics or a validated training simulator. We have not called the API or independently assessed its output.
The stream and controls are specific
The model reference specifies 1280×720 output in a fixed 16:9 aspect ratio at 24 frames per second, with synchronized 48 kHz stereo audio. Inputs include text, a single 16:9 startup image and keyboard or camera user_action events. WASD controls movement; arrow keys control the camera. Mid-session text can request weather, object changes or dialogue.
JavaScript clients can use WebRTC for browser interaction or WebSocket for custom streaming pipelines; the Python async SDK uses WebSocket. Google describes WebRTC delivery as native video and audio tracks, while WebSocket delivers base64 JPEG frames and stereo PCM audio. The guide warns that WebRTC’s 64 KB control-message limit requires compressed reference-image payloads; its recommendation is a JPEG under 48 KB before base64 encoding.
This is not the Gemini Live API, a voice assistant or an offline video-editing endpoint. Genie 4 does not accept live microphone or webcam streams, support function calling or Search grounding, or emit structured outputs. The documentation directs fixed-duration clip generation and editing to Gemini Omni 1.2 Flash instead.
What a minute costs, and how to stop paying
Google’s Genie 4 Standard pricing lists $1.50 per million input tokens for text or reference images, $9.00 per million text output tokens and $17.50 per million video output tokens with audio included. There is no Free tier, Batch API, Flex inference or Priority inference option for this model in the reviewed pricing page.
The documentation bills emitted video at 240 tokens per frame, or 5,760 tokens per second at 24 fps. That gives $0.1008 per second and $6.048 per minute for video with bundled audio, conventionally rounded to about $0.10 per second or $6.05 per minute. These are calculated streaming-output costs, not a guaranteed all-in session invoice; input and any separately billed text output remain additional.
The operational detail worth reading before a trial is session shutdown. Google says closing a browser tab, detaching or leaving Python’s async context does not explicitly complete server-side generation. Call await session.complete() to stop generation and billing, or cancel an active run. Genie 4 does not support session resumption, so treating a disconnected transport as a paused world is also incorrect.
A live world is not a persistent game engine
The model page caps sessions at 120 seconds and says visual coherence and object permanence are strongest in the first 45–90 seconds. The guide warns that stability may degrade after 60–90 seconds and that looking away for eight seconds or turning around can shift or reset objects. It describes a rolling temporal window, not a persistent 3D scene graph.
Google documents roughly 5–9 seconds before the first frame, about 0.2–0.5 seconds for native movement responses under favorable network conditions, and about 1–2 seconds for text-driven changes. These are vendor documentation figures, not AccessAllGPT measurements. Serving locations are the United States, Western Europe and Japan; network distance can increase steering latency and reduce delivered frame rate.
Readable signs, gauges and clocks can deform, and exact object counts can drift. Safety filters may block startup or terminate a stream. Google says generated outputs include SynthID watermarking. None of these statements establishes suitability for a particular commercial project or an independently audited safety guarantee.
Our analysis: the strongest first use is a short visual prototype where an inconsistent landmark is tolerable. A workflow that depends on stable geometry, exact counts, reliable physics or a resumable environment needs a different acceptance bar. A higher frame rate or convincing demo does not remove those requirements.
What builders should try next
For a first evaluation, use a synthetic landscape with a bridge and two distinct landmarks, and a short fixed run budget. Walk away from one landmark, turn back and record whether it stays put; trigger one weather change and measure the response in your own region. Check audio synchronization and separately verify that the server acknowledges completion when the user stops. This is a proposed test, not one we performed.
Keep the production API key on the server and follow Google’s ephemeral-session-token browser pattern. Start with a small number of paid sessions, inspect the actual usage bill, and confirm current account quotas before building a public demo. Compare against an offline video model if users only need a finished clip: live steering is the reason to pay for this interface.
A short Genie 4 pilot
Suggested acceptance checks, not hands-on results.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Choose a brief steerable visual experience. Reject requirements for persistent geometry, exact counts, session resumption or live microphone input.
Measure first frame, keyboard response, text-edit delay, audio sync and landmark consistency from your intended user region.
Use a short fixed budget, complete or cancel explicitly, confirm the terminal event and inspect metered usage before increasing concurrency.
Primary sources
Browse the publication-wide evidence index →
- Genie 4 is now available on the Gemini APIGoogle · Reviewed: November 19 date; Interactions API and AI Studio availability; text/image seeding, 720p at 24fps, native audio, controls and SynthID · Retrieved · Supports: First-party announcement of developer access and vendor-described capabilities; not independent performance evidence.
- Interactive world generation with Genie 4Google AI Studio · Reviewed: Paid-key requirement; SDK transport examples; startup and steering timing; control events; complete, cancel and detach; limitations · Retrieved · Supports: Documented integration and billing-termination behavior, supported transports, timing estimates and limitations; examples were not executed.
- Genie 4 model referenceGoogle AI Studio · Reviewed: Model code; stream/session limits; supported inputs; capabilities; consumption options; regions · Retrieved · Supports: genie-4 identifier, fixed output format, 120-second maximum, media specifications and unsupported features.
- Gemini API pricing: Genie 4Google AI Studio · Reviewed: Genie 4 Standard input/output rates, billing footnote, audio bundling and unsupported pricing tiers · Retrieved · Supports: Published USD token prices and per-frame billing basis; not an independently verified invoice.
- Genie 4 is now available in Project GenieGoogle · Reviewed: November 18 date; eligible Ultra subscriber rollout; Pro waitlist; model limitations · Retrieved · Supports: Consumer product rollout context, distinct from November 19 developer API availability.
Limitations
Public-source briefing based on Google-controlled announcements and documentation. No paid API call, authenticated AI Studio access, generated-world test, independent benchmark, latency measurement, quota verification or invoice audit. Timing, quality and availability are vendor statements; calculated streaming costs exclude applicable input and text-output charges. Documentation can change after retrieval.
Disclosures
Independent public-source reporting and AccessAllGPT analysis, not sponsored testing. Google did not review or endorse this briefing. AccessAllGPT Research is operated by NeuralArc; publication-wide relationships are described on the disclosures page.



