Google brings Genie 4 to the Gemini API at about $6 per minute

Google has opened Genie 4, its real-time interactive world model, to developers through the Gemini API and Google AI Studio. The model, listed under the code genie-4, generates a continuous 720p world at 24 frames per second with synchronized stereo audio, and it responds to keyboard movement, camera controls and text instructions while the stream is running. Google announced the developer release on November 19, 2026, one day after it brought Genie 4 to Project Genie for Google AI Ultra subscribers.
The official Gemini API pricing page lists Genie 4 at $1.50 per 1M input tokens for text and reference images, $9.00 per 1M output tokens for text and $17.50 per 1M output tokens for video, with audio included in the video rate. Because Google bills 240 video output tokens per frame, a running world consumes 5,760 tokens per second, which Google itself converts into an effective rate of about $0.10 per second, or roughly $6.05 per minute. Prices as listed on the official pricing page on November 20, 2026.
Access requires a paid-tier Gemini API key. Google's pricing page states that Genie 4 does not support the Free tier, the Batch API, Flex inference or Priority inference, and context caching is listed as not available, so the standard rate is the only rate. A single session can run for up to 120 seconds, which caps the video cost of one world at about $12.10.
··········
GENIE 4 GEMINI API PRICING: THE OFFICIAL RATES.
Genie 4 is priced on the same token logic as Google's video models, with one difference that shapes every budget: the model streams continuously, and billing runs for as long as the session stays open. The text and image inputs that set up a world cost very little next to the video stream, so the per-second video rate is the figure developers will plan around.
........
Genie 4 item (Standard, paid tier) | Official price | Notes |
Input: text and reference image | $1.50 per 1M tokens | World prompt, mid-session text, one 16:9 image |
Output: text | $9.00 per 1M tokens | Text emitted by the session |
Output: video, audio included | $17.50 per 1M tokens | 240 tokens per frame, 5,760 tokens per second |
Effective video rate | about $0.10 per second | About $6.05 per minute, per Google |
Context caching | Not available | No cached-input price |
Batch, Flex, Priority | Not supported | No discounted tier |
Free tier | Not supported | Paid-tier API key required |
........
Google's footnote on the pricing page spells out the arithmetic: billing is based on emitted video output tokens at 240 tokens per frame, the 48 kHz stereo audio is bundled into the video token rate, and the result is an effective rate of approximately $0.10 per second. The same note tells developers to always call complete() to end an active session and stop billing.
··········
GENIE 4 IN THE INTERACTIONS API: HOW A SESSION WORKS.
Genie 4 does not run through the Live API or the classic generateContent endpoint. Google's documentation places it on the Interactions API in v1beta, where an application creates an interaction on its server and then connects to the stream over WebRTC for browsers or WebSocket for Python and Node.js. The first call needs only the model name, as in Google's own REST example.
curl -s -X POST \
"https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "x-goog-api-key: ${GEMINI_API_KEY}" \
-H "Content-Type: application/json" \
-d '{"model": "genie-4"}'After the connection opens, the client sends one world_setup event with a text prompt, an optional 16:9 reference image or both, and then streams user_action events for movement and camera control or input events with new text instructions. Movement uses the W, A, S and D keys and the camera uses the arrow keys; held keys keep acting until a key-up event arrives. Google notes that the E key and the space bar exist in the schema without built-in behavior, and it recommends mapping custom actions to text inputs instead. For browser apps, the documentation tells developers to mint a session-scoped token on the server rather than expose the API key.
The output is 1280x720 video at a fixed 24 frames per second in 16:9 only, with 48 kHz stereo audio delivered as 16-bit PCM over WebSocket or as an Opus track over WebRTC. Initial generation takes about 5 to 9 seconds, native movement shows up after roughly 0.2 to 0.5 seconds, and text instructions take about 1 to 2 seconds to appear in the world. Google lists the model as stable under the version name genie-4.
··········
GENIE 4 SESSION LIMITS, REGIONS AND MISSING FEATURES.
The hard cap for a Genie 4 session is 120 seconds, and Google's model page says visual coherence and object permanence are strongest during the first 45 to 90 seconds, with the developer guide warning that stability may degrade after 60 to 90 seconds. Session resumption is not supported, so a world that ends cannot be picked up again. The documentation also lists drift in readable text, clocks and exact object counts, and objects that may shift or reset after they stay out of view for eight seconds or more.
Serving runs from data centers in the United States, Western Europe and Japan. Google says round-trip latency under 150 milliseconds is best reached from North America, Western and Central Europe and East Asia, and users elsewhere can expect slower steering and possibly lower frame rates. Concurrency limits are linked from the documentation but no Genie 4 figures appeared on the public rate-limits page when it was checked on November 20.
The feature list is narrow by design. Genie 4 does not support function calling, search or Maps grounding, structured outputs, code execution, file search, URL context or context caching, and it accepts no live microphone, webcam or user audio input. Prompts or images showing identifiable real people, public figures or copyrighted brands may be blocked or may end a stream, and every output carries Google's SynthID watermark.
··········
[[ADPLUS_300x250_1]]
WHAT GENIE 4 COSTS FOR TYPICAL WORKLOADS.
The cost model is simple once the token rate is fixed: each second of world costs 5,760 tokens at $17.50 per 1M, or $0.1008. Input costs barely register, since even 10,000 input tokens of prompts and steering text cost $0.015. The examples below use only the official rates and leave out text output, which most world sessions produce little of.
........
Genie 4 workload (assumptions) | Video tokens | Video cost |
One 30-second world | 172,800 | $3.02 |
One 60-second world | 345,600 | $6.05 |
One 120-second world (session cap) | 691,200 | $12.10 |
Game studio prototyping: 20 designers x 8 sessions x 60 s per workday | 55.3M per day | $967.68 per day, $21,288.96 per 22 workdays |
Public web demo: 1,000 visitors x 30 s per day | 172.8M per day | $3,024 per day, $90,720 per 30 days |
........
The figures show where Genie 4 fits. Internal prototyping by a small team, where a designer opens a world, tests a mechanic or a scene and closes it, stays in the range of a few hundred dollars a day. A consumer-facing demo is another matter, because every visitor's session is billed at the full streaming rate and there is no batch or cached path to bring the price down. A demo that lets 1,000 people play for 30 seconds a day runs above $90,000 a month.
Session hygiene has a direct price. Google's documentation says that closing a browser tab or detaching the client does not stop generation on the server; only session.complete() or session.cancel() ends billing. A 30-second session that is abandoned without a call to complete() keeps running until the 120-second cap, adding about $9.07 to a session that should have cost $3.02.
··········
GENIE 4 VS GEMINI OMNI FLASH AND PROJECT GENIE.
The closest reference points on Google's own price list are the Gemini Omni Flash video models, which generate clips rather than interactive worlds, so the comparison concerns the price of each second of 720p video and not equivalent products. Gemini Omni 1.2 Flash lists video output at $12.00 per 1M tokens, which Google puts at about $0.07 per second, while Gemini Omni 1.1 Flash lists $17.50 per 1M tokens, about $0.10 per second, and Google's pricing page says Omni 1.1 routes to the 1.2 rates after December 3, 2026.
........
Google model (Gemini API) | Video output price | Effective rate (Google) | Notes |
Genie 4 (interactive world, 720p, 24 fps) | $17.50 per 1M tokens | about $0.10 per second | No batch, no caching |
Gemini Omni 1.2 Flash (video generation) | $12.00 per 1M tokens | about $0.07 per second at 720p | Announced November 5, 2026 |
Gemini Omni 1.1 Flash (video generation) | $17.50 per 1M tokens | about $0.10 per second | Routes to 1.2 rates after Dec 3 |
........
A 60-second Genie 4 world therefore costs about $6.05, the same as a minute of Omni 1.1 output and about 44% more than the roughly $4.20 that a minute of Omni 1.2 output costs at Google's approximate rate. The premium buys interactivity: an Omni clip is generated and delivered, while a Genie 4 world is steered live by the user for the whole session. We covered the Omni price cut in our report on Gemini Omni 1.2 Flash pricing.
Consumers reach the same model through a different door. On November 18 Google brought Genie 4 to Project Genie for eligible Google AI Ultra subscribers and opened a waitlist for Google AI Pro, as described in our article on Genie 4 in Project Genie. Subscribers pay through their plan, while developers pay per second of stream, so an app built on the API carries the full metered cost of every user session. Developers planning a budget on top of other Gemini models can check our guide to Google AI Studio pricing, free access and API costs.
··········
GENIE 4 ON THE GEMINI API: CONFIRMED FACTS AND OPEN QUESTIONS.
The confirmed facts come from Google's announcement, model page, developer guide and pricing page. Genie 4 is available as a stable genie-4 model on the Interactions API to paid-tier keys, with official rates of $1.50 per 1M input tokens, $9.00 per 1M text output tokens and $17.50 per 1M video output tokens, an effective cost of about $0.10 per second, a 120-second session cap, 720p output at 24 frames per second and serving from the United States, Western Europe and Japan. Batch, Flex, Priority, caching and the Free tier are all excluded.
Several points remain open. Google has not published Genie 4 concurrency figures on its public rate-limits page, has not given a Vertex AI listing in the developer announcement, and has not said whether longer sessions, resumption or higher resolutions are planned. The research preview that Google DeepMind described in October, covered in our earlier report on the Genie 4 world model, has now become a metered product whose cost is set almost entirely by how many seconds of world an application streams and how reliably it closes each session.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




