
Genie 4 API Is GA: About $6 a Minute, Capped at 120 Seconds
Genie 4 API Is GA: About $6 a Minute, Capped at 120 Seconds
Google flipped the switch on the Genie 4 API on November 19. The Genie 4 world model, genie-4, is now generally available in the Gemini API and AI Studio, one day after Project Genie reached AI Ultra subscribers.
You send a prompt or an image. You get back a world you can walk around in.
It's not a new model. DeepMind announced Genie 4 back on October 27, and we covered the model's two-minute worlds then. What's new is that developers can call it, pay for it, and stream it into their own apps.
At roughly $0.10 per second of output, that's about $6 a minute, or $360 for an hour of continuous worlds. And no single world lasts longer than two minutes.
Genie 4 API Key Features and Technical Breakdown
Genie 4 generates an interactive, continuous world in real time. You seed it with a text prompt or a reference image. Then you steer it while it runs.
Here's what Google's Genie docs confirm:
- Output: 1280x720 at 24 fps, 16:9 only, with 48 kHz stereo audio.
- Access: model ID
genie-4, on thev1betaInteractions API. Nothing else. - Transport: WebRTC is recommended for browsers. WebSocket is the server-side option.
- Controls: WASD moves, arrow keys turn the camera. Both are sent as
user_actionkey events. You can also send text mid-session to change what's happening. - Watermarking: every output carries a SynthID watermark.
- Browser auth: use short-lived
auth_tokens so your API key never ships to the client.
That control model matters. This isn't a video generator where you wait for a clip. It's a live stream that reacts to keypresses, which is closer to a game engine than to a text-to-video model.
Genie 4 Pricing in the Gemini API: The Per-Second Math
Google's pricing page lists Genie 4 at:
| Item | Price |
|---|---|
| Input | $1.50 per 1M tokens |
| Text output | $9.00 per 1M tokens |
| Video output (audio included) | $17.50 per 1M tokens |
| Effective video rate | approximately $0.10 per second |
It's paid tier only. There's no free tier, and no Batch, Flex or Priority options.
Token prices are hard to picture for a live stream, so here's the math in seconds. We haven't billed a production workload yet. These figures are our own arithmetic on Google's published rates, so check them against your first invoice:
| Session length | Approx. video output cost |
|---|---|
| 30 seconds | $3 |
| 60 seconds | $6 |
| 90 seconds | $9 |
| 120 seconds (hard cap) | $12 |
A few more numbers worth keeping handy:
- Implied token rate: $0.10 per second at $17.50 per million works out to roughly 5,700 video tokens per second of output. That's useful if you're estimating spend from token logs.
- A small demo: 1,000 users each playing one full two-minute session costs around $12,000.
One caution on those per-token numbers. Outside pricing write-ups list the same rates for Gemini Omni 1.1 Flash. So don't assume the rates were set specially for Genie.
The 120-Second Wall
This is the limit that'll shape your product. Google's Genie docs say quality may start to degrade after 60 to 90 seconds, and every session has a hard cap of 120 seconds.
There's also no session resumption. When the clock runs out, that world is gone. You can't save it and pick it back up later.
Other gaps from the docs:
- No microphone or webcam input.
- No function calling, so the world can't trigger your backend tools mid-session.
- Reference images should stay under 48 KB, which means heavily compressed thumbnails, not concept art.
Real-World Use Cases That Fit Those Limits
Two minutes is short. Build around it, not against it.
Interactive previews. A travel site, real estate listing or game studio can let someone step into a concept for 30 to 60 seconds. That stays inside the quality window and costs about $3 to $6 per view.
Pitch and prototyping tools. Level designers and filmmakers can sketch a mood, walk it for a minute, then throw it away. Paying $6 to test a vibe is cheap next to manual blockout work.
Paid, gated experiences. At $6 a minute, an open free demo burns money. A per-session credit system or a premium tier is the realistic model.
Not a fit: anything persistent. Open-world games, training environments that need hours of continuity, or apps where users expect to come back to their world all run into the 120-second cap and missing resumption.
Hard Limitations and Flaws
Let's be blunt about the tradeoffs.
It's expensive for consumer apps. $6 a minute is fine for a B2B demo. It's brutal for a free app with viral traffic.
Put rate limits and per-user caps in place before you launch anything public.
Sessions are disposable. No resumption and a two-minute ceiling mean you can't build lasting state on top of Genie alone.
It's sealed off from your stack. No function calling and no mic or camera input means your app can steer the world, but the world can't call back into your app.
Quality drifts before the cap. Google says degradation can start at 60 to 90 seconds. Your users may see it before they hit the limit, so test longer sessions yourself.
Interactions API only. It runs on the v1beta Interactions API, so plan for that integration path.
Verdict: Developer Action Plan
The Genie 4 API is real, it's GA, and it lets developers stream a steerable generated world into their own product. It just isn't cheap or persistent yet.
Here's how to use Genie 4 without getting burned:
- Prototype on the Interactions API with WebRTC in the browser and short-lived auth tokens from day one.
- Design for 60 seconds, not 120. Treat the second minute as a buffer, not a feature.
- Budget in seconds. Use roughly $0.10 per second of output and cap sessions per user.
- Keep reference images tiny, under 48 KB.
- Wait on anything persistent until Google adds session resumption or longer sessions.
Short, paid, throwaway worlds? Ship it this week. Anything that needs to remember the player? Not yet.
Tools in this post
Written by
Co-founder of Tool Directory AI. Tests and reviews the AI tools listed here.


