Genie 4 (models/genie-4) is Google DeepMind's real-time interactive world generation model. From a text prompt, a single 16:9 reference image, or both, it builds a world you can walk through with WASD or the arrow keys, streaming 720p video at 24 fps with synchronized 48 kHz stereo audio. It became generally available on the Gemini API (Interactions API) on November 19, 2026, priced at roughly $0.10 per second of output (about $6.05 per minute).
This guide summarizes the specs, pricing, and limits published in Google's official documentation (ai.google.dev) and works out what real-world use actually costs.
What Genie 4 does (and how it differs from video generation)
A typical video model takes a prompt, makes you wait, and returns a fixed-length clip. Genie 4 instead renders frames on the fly and streams them continuously, changing the next frame in response to your input — closer to a game engine than to a clip generator.
Item | Genie 4 | Gemini Omni 1.2 Flash (for comparison) |
|---|---|---|
Purpose | Steerable, real-time world generation | Fixed-duration clips, multi-turn video editing, upscaling to 4K |
Inputs | Text, one 16:9 reference image, keyboard/camera controls, mid-session text events | Text, images, video |
Output | 1280x720 (720p, 16:9) at 24 fps + 48 kHz stereo audio | Video clips |
Length | Up to 120 seconds per session | Per clip |
Transport | WebRTC (browsers) or WebSocket | Standard requests |
Google's own docs point to Gemini Omni 1.2 Flash for offline clip generation and editing.
Controls: keys and mid-session text
There are two ways to steer. Movement keys (WASD and arrows) move you while held. Mid-session text events change the world — the official samples bind keys to prompts such as "make heavy snow fall across the painted village." According to the docs, the first frame takes about 5–9 seconds to arrive, and text-driven changes appear in about 1–2 seconds.
Quality limits to know
- Coherence is strongest in the first 45–90 seconds. Sessions can run to 120 seconds, but visual coherence and object permanence are best early on.
- No persistent 3D mesh. Turn 180 degrees and back, and background buildings may rearrange; Genie 4 keeps a rolling visual horizon rather than a stored 3D scene.
- US-only serving. Round-trip latency above 150 ms noticeably delays keyboard steering, so test from your users' region first.
Genie 4 API pricing (as of November 19, 2026)
From the Gemini API pricing page (USD per 1M tokens):
Item | Free tier | Paid tier |
|---|---|---|
Input (text / reference image) | Not available | $1.50 |
Output (text) | Not available | $9.00 |
Output (video, audio included) | Not available | $17.50 |
Context caching | Not available | Not available |
Used to improve Google products | — | No |
Video is billed at 240 tokens per frame, or 5,760 tokens per second at 24 fps. 5,760 × $17.50 / 1M = about $0.1008 per second, or about $6.05 per minute — matching Google's own figure. There is no free tier, and Batch API, Flex, and Priority inference are not supported.
What does it cost? Worked examples
Video output is almost the entire bill, so output seconds are all you need for an estimate (input tokens are negligible).
Scenario | Output time | Estimated cost |
|---|---|---|
A quick 30-second test | 30 s | ~$3.02 |
One full-length session | 120 s | ~$12.10 |
Tier 1 daily cap (50 sessions × 120 s) | 100 min | ~$605/day |
One trade-show kiosk running 8 hours non-stop | 480 min | ~$2,903/day |
Back-to-back 120-second sessions for 8 hours means 240 sessions — beyond Tier 1's 50 per day, so you would need Tier 3 (1,000 per day). Capping each visitor at 30–45 seconds and pausing between visitors cuts the cost several-fold, and it also matches the docs' note that the first 45–90 seconds look best.
The billing trap: always call complete()
The docs repeat this warning: closing a browser tab or leaving a code block detaches the transport but does not immediately stop server-side generation. Always call complete() to end the session. At $0.10 per second, an orphaned session adds up quickly.
Access requirements and rate limits
Genie 4 requires a paid-tier Gemini API key and connects through the Interactions API on v1beta. Rate limits are set by concurrent sessions and requests per day:
Tier | Concurrent sessions | Requests per day |
|---|---|---|
Free | Not supported | Not supported |
Tier 1 | 2 | 50 |
Tier 2 | 5 | 250 |
Tier 3 | 10 | 1,000 |
In-session key presses and text events don't count as separate requests. For browser apps, the recommended pattern is to mint a single-session ephemeral token on your server so the API key never reaches the client. To try it first, use the Live playground in Google AI Studio (aistudio.google.com/live?model=genie-4); a paid tier is still required.
From announcement to GA
Google DeepMind announced Genie 4 in late October 2026, starting with a limited group of trusted testers, with Google AI Ultra subscribers (via Project Genie) and API early-access partners to follow, according to coverage at the time. The November 19 Gemini API release notes are the first to publish general availability and pricing.
The same update added Tool retrieval (Preview) to the Interactions API, letting Gemini search a tool library and load function schemas on demand. It works with Gemini 3.6 and later, including all Gemini 4 models (Genie 4 itself does not support function calling). For Gemini 4 pricing, see Gemini 4 Hydrogen pricing and free tier.
Where it fits in real work
Given 720p output, a 120-second cap, and coherence concentrated early in the session, Genie 4 is a tool for building experiences on the spot, not for finished video production.
Good fit | Poor fit |
|---|---|
Walk-in experiences at trade shows and stores | Ad footage that must be reproduced exactly |
Mood prototypes for venues, destinations, or story worlds | Property tours that need accurate dimensions and layout |
Early concepting for games and picture books | Long-form or 4K final video (use Omni or traditional production) |
If you're planning to build something like this into a product or event, Mihata develops custom AI systems for businesses — feel free to take a look.
FAQ
FAQ
What is Genie 4?
It is Google DeepMind's real-time interactive world generation model. From text or a 16:9 image it generates a world and streams 720p, 24 fps video with 48 kHz stereo audio that responds to keyboard controls.
How much does Genie 4 cost?
Video output (audio included) is $17.50 per 1M tokens, which works out to about $0.10 per second or about $6.05 per minute. Input is $1.50 per 1M tokens. There is no free tier.
Is Genie 4 free to use?
No. It is available only on the paid tier of the Gemini API. The Free tier, Batch API, Flex, and Priority inference are not supported.
How long can a session last?
Up to 120 seconds. The docs note that visual coherence and object permanence are strongest during the first 45 to 90 seconds.
Can I use it outside the US?
Yes, with a paid API key, but Genie 4 is served from US data centers only, and round-trip latency above 150 ms noticeably delays keyboard steering.
Summary
Genie 4 is Google's steerable, real-time world generation model, generally available on the Gemini API since November 19, 2026. It costs about $0.10 per second of output (about $6.05 per minute), sessions are capped at 120 seconds, and it runs on the paid tier from US data centers only. Because cost tracks output seconds, decide session length per user up front — and build complete() into your very first prototype.