Google Opens Genie 4 World Generation to Gemini API Developers
Google opened Genie 4 to Gemini API developers on November 19, 2026: 720p interactive worlds at 24 fps for roughly $6.05 per minute, capped at 120 seconds with no persistent 3D scene.

Why it matters
- Google made Genie 4 available via the Gemini API and Google AI Studio on November 19, 2026, one day after Project Genie opened it to AI Ultra subscribers.
- The model generates 720p video at 24 fps with native audio; effective video output pricing is roughly $0.10 per second, or $6.05 per minute.
- Sessions cap at 120 seconds, with visual coherence strongest in the first 45–90 seconds and degradation after 60–90 seconds.
- The model has no persistent 3D scene: objects can reset after eight seconds out of view or a 360-degree turn.
- Tier 3 projects get at most 10 concurrent sessions and 1,000 requests per day; sessions are served only from US data centers.
Google made Genie 4 available through the Gemini API and Google AI Studio on November 19, 2026, opening its interactive world-generation model to any developer with a paid-tier API key. The model generates continuous 720p video at 24 frames per second with synchronized native audio, responding to keyboard controls and live text prompts. The catch sits in the fine print: sessions cap at 120 seconds, and no world survives past the moment a developer closes it.
The release positions Google in a nascent market for programmable, on-demand simulation environments — a category with obvious appeal for robotics training and AI agent experimentation, where synthetic worlds are cheaper than physical ones. But the constraints matter as much as the capability.
How the API actually works
The developer release uses the model identifier models/genie-4 on the v1beta Interactions API. A world can be seeded with text, a single reference image, or both. WASD and arrow keys control movement and camera orientation, while text events modify the scene or supply character dialogue. Google's API documentation explicitly distinguishes this responsive frame rendering from offline generation of a fixed video clip — the model renders in reaction to input, not as a pre-baked asset.
Building a session follows a defined sequence:
- Create an interaction.
- Connect over WebRTC or WebSocket.
- Send one
world_setupevent before steering withuser_actionor text input events.
Google recommends WebRTC for interactive browser applications. WebSocket supports Python and Node.js workflows, frame capture, and custom pipelines. Browser apps can use an ephemeral token scoped to a single interaction, keeping the long-lived API key on the backend — a sensible security posture for client-facing code.
What does the pricing look like?
Genie 4 requires a paid-tier Gemini API key and is not offered on the free tier. Standard pricing runs $1.50 per million tokens for text or reference-image input, $9 per million for text output, and $17.50 per million for video output with stereo audio included. Google lists no volume discounts or alternative serving modes for this model.
At 240 video output tokens per frame — 5,760 tokens per second — the listed effective rates work out to approximately $0.10 per second and $6.05 per minute. Batch API, Flex inference, and Priority inference are all unsupported.
One billing behavior deserves specific attention. Closing a browser tab or detaching the local transport does not immediately stop server generation. Developers must call complete() to end generation and billing, or cancel() to abort server-side generation outright. Explicit shutdown is part of cost control, not an optional courtesy — leave a session dangling and the meter keeps running.
Why does everything stop at 120 seconds?
Across its model documentation and implementation guide, Google caps sessions at 120 seconds and identifies the first 45 to 90 seconds as the strongest period for visual coherence and object permanence. The guide warns that visual stability may degrade after 60 to 90 seconds. In plain terms: half the advertised session length is already past the model's comfort zone.
The technical explanation is structural. The guide describes generation over a rolling temporal window without a permanent 3D scene. Look away from an object for eight seconds or more, or turn 360 degrees, and landmarks, doors, and buttons can move or reset. For robotics or agent experiments, this means the generated environment cannot be assumed to provide a stable long-term map — a significant limitation for exactly the use cases Google promotes.
The model's feature gaps compound the continuity problem:
- Fixed 16:9 aspect ratio, with no portrait mode or configurable frame rate.
- Sessions served only from US data centers.
- No live microphone or webcam stream input.
- No function calling, Google Search grounding, structured outputs, or ability to resume a completed interaction.
Latency figures come from Google's implementation guide. Native keyboard actions get a visual response of about 0.2 to 0.5 seconds at low network latency; mid-session text changes take roughly one to two seconds. Round-trip network latency above 150 milliseconds noticeably delays keyboard steering. The US-only serving requirement means developers and users elsewhere will routinely sit on the wrong side of that threshold.
Who gets how many sessions?
Rate limits scale with a paid project's usage tier, and the top tier is modest:
| Usage tier | Concurrent active sessions | Requests per day |
|---|---|---|
| Tier 1 | 2 | 50 |
| Tier 2 | 5 | 250 |
| Tier 3 | 10 | 1,000 |
One carve-out eases the arithmetic: real-time control messages within a session do not consume separate per-minute or per-day request quotas. Even so, ten concurrent sessions at Tier 3 signals that Google is sizing this for experimentation, not production traffic.
Where does this fit in Google's rollout?
The API announcement landed one day after the Project Genie update on November 18, 2026, which gave eligible Google AI Ultra subscribers their first access to Genie 4 and allowed Google AI Pro members to join a waitlist. That subscriber release emphasized browser-based creation, a Project Genie agent, and a community gallery. The API instead adds developer sessions, documentation, paid-tier pricing, and usage quotas.
Google lists prototyping interactive experiences, visualizing spatial concepts, and training robotics and AI agents as possible uses. The company says every output carries imperceptible SynthID watermarking (Google's embedded content provenance marker), which addresses provenance for generated video but does nothing about the model's spatial reliability limits.
The API makes Genie 4 a programmable environment for brief, controlled simulations rather than a hosted persistent world. Its central constraint is continuity: applications can generate responsive scenes on demand, but the current design does not provide a world they can retain and revisit across completed sessions. Whether Google closes that gap — persistent state, longer sessions, stable spatial memory — will determine whether Genie 4 grows beyond a simulation sandbox into infrastructure for the robotics and agent pipelines Google says it wants to serve.
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- Google DeepMind ships Genie 4: photorealistic worlds, audio, and real-time steering
- Genie 4 reaches Project Genie, bringing live world-editing to AI Ultra users
- Google DeepMind's Genie 4 simulates worlds in real time — with sound and talking characters
- Google Opens Project Genie to AI Ultra Subscribers in the U.S.
- Google Ships Gemini 4 Carbon and Genie 4, Its Second Frontier Release in Two Weeks