DeepMind Genie 4 API Pricing & Limits Review: $17.50/1M (~$6.05/min)

Google DeepMind’s Genie 4 (models/genie-4) streams real-time, keyboard-steerable 720p 24 fps interactive worlds with synchronized 48 kHz stereo audio over the Gemini v1beta Interactions API. There is no Free tier: Paid Standard rates are $1.50 per 1M input tokens and $17.50 per 1M video+audio output tokens—an effective streaming cost of ~$0.10 per second (~$6.05 per minute), capped at 120 seconds per session.

Key Genie 4 API takeaways

1 · Streaming Token Rate240 video tokens per frame at 24 fps equals 5,760 tokens/sec (~$0.1008/sec or ~$6.05/min), with 48 kHz stereo audio bundled.
2 · Strict 120s Session CapMax session duration is 120 seconds (strongest visual coherence in the first 45–90 seconds); always call complete() to stop billing.
3 · Paid-Only & Regional RTTNo Free, Batch, Flex, Priority, or caching tiers. Served in US, Western Europe, and Japan (<150 ms RTT recommended).

Official Genie 4 API pricing & technical limits

Published figures from Google Gemini API Pricing and Genie 4 Model Documentation (verified November 20, 2026).

ParameterGenie 4 (models/genie-4)Notes & Billing Impact
Free Tier / Batch / Flex / PriorityNot supportedRequires a Paid-tier Gemini API billing key; context caching is also unavailable.
Input Price (Paid)$1.50 / 1M tokensCovers text world_setup prompts, mid-session text steering, and a single 16:9 seed image.
Text Output Price (Paid)$9.00 / 1M tokensApplies to text event outputs; video+audio stream uses the separate rate below.
Video + Audio Output (Paid)$17.50 / 1M tokens240 tokens/frame at 24 fps = 5,760 tokens/sec (~$0.10/sec; ~$6.05/min); 48 kHz stereo audio bundled.
Output Stream Specs1280×720 (720p, 16:9) · 24 fps48 kHz stereo audio (16-bit PCM via WebSocket; Opus via WebRTC).
Max Session Duration120 secondsCoherence and object permanence hold best in the first 45–90 seconds.
WebRTC Reference Image CapKeep JPEG < 48 KBRequired so base64 JSON in world_setup stays within WebRTC’s 64 KB SCTP data-channel limit.
Serving RegionsUS, Western Europe, JapanRecommend <150 ms RTT; higher latency delays WASD/arrow steering and may drop frame rate.
Paid Data Training PolicyNoPaid-tier Gemini API inputs and outputs are not used to improve Google products.

Session cost math & engineering gotchas

1. Real-world streaming cost per session: At 240 tokens/frame × 24 fps = 5,760 output tokens/sec and $17.50 / 1M, pure video+audio output costs $0.1008 per second (prompt and reference-image input tokens at $1.50/1M add fractions of a cent per session):

Session LengthVideo Output TokensEstimated Output CostRecommended Use Case
15 seconds86,400 tokens~$1.51Interactive preview / automated QA test run
60 seconds345,600 tokens~$6.05Single scene exploration inside optimal 45–90s coherence window
120 seconds (Max)691,200 tokens~$12.10Full continuous session ceiling before mandatory disconnect

2. Always call session.complete(): Unlike request/response endpoints that stop after a fixed token count, Genie 4 generates continuously on the server until explicitly closed or hitting the 120-second ceiling. Orphaned browser tabs or uncaught WebSocket disconnects keep accruing ~$0.10/sec until cleanup; mint single-session ephemeral auth tokens server-side via interactions.authTokens.create() and trigger complete() on user exit or inactivity timers.

3. Genie 4 API vs Project Genie vs Gemini Omni 1.2 Flash: For non-interactive linear clips, Google’s own Genie 4 spec directs builders to Gemini Omni 1.2 Flash (offline fixed-duration clips up to 15s, multi-turn editing, and 4K upsampling). Consumer access via Project Genie runs inside Google AI subscriptions (Google AI Ultra priority access; Pro waitlist) rather than direct token billing.

When to choose Genie 4 API

Pick models/genie-4 when: you are building browser-based playable worlds, spatial simulations, robotics/agent environments, or interactive prototypes where users steer a 720p world in real time using WASD/arrows and live text triggers.

Skip Genie 4 API when: you need 1080p/4K video export, deterministic offline renders longer than two minutes, or sub-$0.50/minute unit economics—use rendered video models or Gemini 4 Neon / Hydrogen / Carbon text/multimodal APIs instead.

Related guides & official docs: Gemini 4 Neon vs Hydrogen vs Carbon Pricing · How to Access Gemini 4 Carbon · Best AI Video Generators 2026 · Interactive World Generation with Genie 4 (Google AI Studio) · Try Genie 4 in Google AI Studio Live.

Frequently asked questions

How much does the DeepMind Genie 4 API cost per minute?

At $17.50 per 1M video+audio output tokens (240 tokens/frame at 24 fps = 5,760 tokens/sec), Genie 4 costs roughly $0.10 per second or ~$6.05 per minute of active streaming, plus $1.50 per 1M input tokens.

Does Genie 4 have a Free API tier or Batch discount?

No. Genie 4 is available only on the Paid tier of the Gemini API and does not support the Free tier, Batch API, Flex inference, Priority inference, or context caching.

What is the maximum session length and reference image size?

Each interactive session caps at 120 seconds (best visual coherence in the first 45–90 seconds). When using WebRTC with a 16:9 seed image in world_setup, keep the compressed JPEG under 48 KB so the JSON message stays under the 64 KB SCTP limit.

How do I prevent runaway Genie 4 API charges?

Mint single-session ephemeral tokens with interactions.authTokens.create(), enforce client-side idle timeouts, and always call session.complete() when the user pauses or leaves the stream.

Specifications and USD rates reflect official Google AI Studio documentation on November 20, 2026; verify live tier concurrency quotas and regional endpoint latency before production deployment.