500xEngineer.

All postsSHORT4 min read

Genie 4: Playable AI Worlds for About 10 Cents a Second

Google opened Genie 4 in the Gemini API on Nov 19, 2026. How controls, pricing, session limits and the complete() billing stop work.

Google now sells playable AI worlds for about 10 cents a second. In your own API app, closing the tab can leave the meter running. On Nov 19, 2026, Google opened Genie 4 to developers in the Gemini API. You seed a world, steer the streaming video, and explicitly end the session to stop billing.

Video not loading? Watch the Short on YouTube

What Google launched

The Gemini API changelog labels models/genie-4 generally available on the Interactions API. Its API version is v1beta. It's also available in Google AI Studio's Live playground with a paid-tier API key.

Logan Kilpatrick's launch post called it “our first world model for developers to build with”. Google recommends it for steerable worlds, playable visual prototypes and interactive spatial exploration.

Keep the launch dates separate. On Nov 18, Google began rolling Genie 4 out to eligible Google AI Ultra subscribers through its consumer app, Project Genie. The Gemini API developer launch followed on Nov 19, 2026.

How a session works

The session docs describe the lifecycle:

  • Create: create an interaction for models/genie-4 on v1beta.
  • Connect: open its live session over WebRTC or WebSockets.
  • Seed: provide text, a 16:9 reference image, or both. Initial generation takes about 5-9 seconds before the first frame.
  • Steer: use WASD, arrow keys and mid-session text events while video and sound stream.
  • Complete: call complete() when finished to end server generation and stop billing.

The stream is 1280x720 (720p), 16:9, 24 fps, with synchronized 48 kHz stereo audio. The output format is fixed.

In your own API app, closing the tab disconnects without immediately stopping server generation, so generation and billing can continue until the session is completed, cancelled or reaches the 120-second cap. Call complete() in your exit flow to end generation and billing; cancel() aborts generation. Google AI Studio’s playground pauses or ends the stream after a tab is hidden for more than 10 seconds.

What you can control

Keyboard movement takes effect in about 0.2-0.5 seconds. Mid-session text prompts take about 1-2 seconds to appear. You can change the world as you explore it.

The controls guide includes an AI Studio starter world called “Step inside a living oil painting”. Custom keys bind reusable action prompts. Google's example binds E to “Make heavy snow fall across the painted village”.

Add Rule: clauses to steer how objects affect the world. These rules steer probabilistically, so never promise reliable game mechanics. The guide's example is “Pulling the red lever makes the stars outside the window stretch into bright hyperspace streaks”. You can also put text in quotation marks to converse with a character in front of you.

What it costs

Genie 4 requires the paid tier. Google's pricing page lists these rates per 1M tokens:

Token type Price
Input: text or reference image $1.50
Text output $9.00
Video output, audio included $17.50

Video output uses 240 tokens per frame, or 5,760 tokens per second at 24 fps. Google gives an effective rate of approximately $0.10 per second, or ~$6.05 per minute. That's about $6 a minute when you're explaining the cost out loud.

CALC: about $12 of video output per full 2-minute session. That's our arithmetic: 2 x ~$6.05 = $12.10, rounded to about $12. It excludes input and text output costs.

The catch

Google describes these world models as being in the frontier research space. Build around the limits:

  • Session length: 120 seconds is the hard cap. Visual stability may degrade past 60-90 seconds.
  • Memory: looking away for 8+ seconds or turning 360 degrees can shift or reset landmarks, doors and buttons.
  • Format: 16:9 only. Portrait 9:16 and configurable frame rates aren't supported.
  • Content filters: identifiable real people, public figures and copyrighted brand IP can be blocked at startup or terminate an active stream.
  • Detail drift: signs, gauges and clocks deform over time. Exact object counts can drift across camera pans.
  • Inputs and integrations: no microphone or webcam input, tool calling, search grounding, structured outputs or session resumption.
  • Serving regions: data centers in the US, Western Europe and Japan. Connections outside North America, Western and Central Europe and East Asia can have higher steering latency and lower frame rates.
  • Rate limits: Tier 1 allows 2 concurrent sessions and 50 requests per day. Requests are the published daily unit.
  • Billing: always call complete() to end the session. In your own API app, closing the tab can leave generation and billing running until completion, cancellation or the 120-second cap.

On generated content, Google's terms say Google won't claim ownership over what you generate, may generate similar content for others, and leaves responsibility for use with you.

The builder angle: my opinion

I'd offer small game studios a fixed-scope playable pitch: their concept as an interactive world, with a prompt, world rules and custom keys. Run it live in the pitch meeting as the operator.

Make the guard rails part of the deliverable. Plan short runs: stability is best under 60-90 seconds. Include a 2-minute session timer, complete() on exit and a spend cap you build. Tier 1’s spend limit is $10 per rolling 10 minutes where it applies, depending on billing history and account standing. A full 2-minute run costs about $12 in video output (CALC: 2 x ~$6.05), above that allowance. World rules steer probabilistically, so the deliverable is a visual prototype; never promise reliable game mechanics. Use concepts without real people or brand IP. Plan operator-run demos around Tier 1's 2 concurrent sessions, and keep its 50 requests per day limit visible.

This is my opinion, with no income guaranteed.

Daily email

Get each breakdown in your inbox.

One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.

Read past emails on Substack

Sources

All postsWatch on YouTube