sysnote
News

Genie 4 reaches the Gemini API: 120-second worlds at about $0.10 a second

Google made Genie 4 available in the Gemini API on November 19, as models/genie-4 on the Interactions API (v1beta), paid tier only. A day earlier the world model had reached Google AI Ultra subscribers in Project Genie. When DeepMind announced it on October 27, it gave no resolution, frame rate, session length or price. The developer documentation now gives all four, along with a fairly frank list of where the model breaks.

What a session looks like

An application creates an interaction on its server, mints a single-session token for the client, and the client connects over WebRTC (in a browser) or WebSocket. The first event, world_setup, carries a text prompt and optionally one 16:9 reference image; the first frame takes about 5–9 seconds to arrive. After that the model streams 1280×720 video at 24 fps with 48 kHz stereo audio. The player steers it with user_action events, key_w, key_a, key_s and key_d for movement and the arrow keys for the camera, and the application can change the world mid-session with text input events such as "Make heavy snow fall across the painted village".

A session lasts at most 120 seconds, and the model page says visual coherence and object permanence are strongest during the first 45–90 seconds. The limitations section explains why: Genie 4 "simulates frames over a rolling temporal window rather than maintaining a persistent 3D scene graph". An object that has been out of view for 8 seconds or more may have moved or reset when the player looks back. In the prompting guide's espresso-machine example, a mug that has just been filled can be empty again after the camera pans 90–180 degrees away from the counter and returns.

Project Genie's updated Genie 3, by comparison, now runs for up to 3 minutes per generation, up from 60 seconds, according to Google's November 18 post. Google keeps it in the model picker "when you want longer runs".

Keys and text take different paths

Movement and camera keys take effect in about 0.2–0.5 seconds. Text goes through what the documentation calls world planning and shows up on screen in about 1–2 seconds. The planner picks up text only at step boundaries, every 4 frames, so a single message can be missed: Google's sample code sends each one-off prompt 10 times at 150 ms intervals and recommends a 2-second cooldown before sending a conflicting one.

That is also how a developer gets action buttons. key_e and space exist in the user_action enum, but the guide says the model has no trained "interact" or "jump" behavior for them. The suggested pattern is to catch E, Q or R in the client and send a text prompt instead. Cause-and-effect rules can be written into the setup prompt ("Rule: Pulling the red lever makes the stars outside the window stretch into bright hyperspace streaks"), and the guide calls these "probabilistic steering, not a game script". It describes its phrasing rules as empirical workarounds, and they read that way: bind a rule to an object rather than a place, because the model "cannot measure abstract zone boundaries", and put the lever "far away from me", because one within arm's reach often fires during setup before the player has moved.

The planner also edits what it is given. According to the guide, Genie 4 strips or rewrites requests for soundtrack music and for "narrowband drone" sounds such as "engine hum" or "jet roar". If a prompt describes fast motion without saying what moves the camera, the model often invents a boat hull or a vehicle hood partway through. Google's advice is to ask for physical sounds like footsteps, gravel and water, and to put the means of travel in the opening sentence.

In October, DeepMind described a different kind of loop for its demos: Gemini 3.8 Flash observing Genie 4's frames, breaking a goal into steps, checking each step and re-steering the world. Project Genie now offers a version of this, with Gemini watching the screen and triggering events. The API has no built-in equivalent. Genie 4 has no function calling and accepts only text, one startup image and key events, so an application that wants a model watching the scene would have to run that model itself on the video frames and send its decisions back as text events, each subject to the same 1–2 second delay.

What it costs

Billing counts output frames: 240 video tokens per frame, or 5,760 tokens per second, at $17.50 per million tokens with the audio included. That comes to $0.1008 per second, the "approximately $6.05 per minute" on the pricing page, so a full 120-second session costs about $12.10. Text and image input costs $1.50 per million tokens and text output $9.00; the documentation doesn't say how much text output a session produces. There is no free tier, and Batch, Flex and Priority inference are not available. For comparison, offline 720p clips from Gemini Omni 1.2 Flash cost about $0.07 a second when it went GA in November (our report); a second of live, steerable video costs about 40% more.

Billing stops when the client calls complete() or cancel(). Closing the browser tab or leaving a Python async with block only detaches the connection, and the documentation warns that generation and billing continue on the server. It doesn't say how long a detached session keeps running; if it runs to the 120-second cap, an abandoned session costs as much as a finished one.

Rate limits are counted in sessions rather than requests: 2 concurrent sessions and 50 per day at Tier 1, 5 and 250 at Tier 2, 10 and 1,000 at Tier 3. Control messages during a session don't count against them. A public demo on a Tier 1 key can host two players at a time. Genie 4 is served only from US data centers; the documentation says round-trip latency above 150 ms noticeably delays keyboard steering, and its driving example asks for under 100 ms. There are no EU or Asia-Pacific endpoints, and Google suggests running backend relays in North America and testing WebRTC latency from the target region before building anything that depends on quick reactions.

Google still hasn't published a technical report on Genie 4. Every number here, including the 45–90 second coherence window, comes from Google's own documentation, and we haven't seen independent tests of the API yet.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.