AI DevelopmentCost Playbook9 min readPublished

What It Costs to Run a Live World With Google Genie 4

Thirty seconds of steerable video costs about $3.02 in output. Keeping a world open for two minutes costs about $12.10. The important part is what the visitor gets to choose during those seconds.

DA
Digital Applied
PublishedNovember 21, 2026
Read time9 min

Suppose a museum wants to know which archway a visitor will explore when music starts playing in a courtyard. Google Genie 4 can generate that courtyard as the visitor walks through it. At full frame rate, half a minute costs about $3.02 in video output, sound included. If the visitor leaves after thirty seconds and the application lets the world keep running to its two-minute cap, the output could reach about $12.10 for the same visit.

Google opened Genie 4 through the paid Gemini Interactions API on November 19, a day after bringing it to eligible Project Genie subscribers. A developer can now put the controls in a prototype and pay for the stream. The museum can reserve that spend for its interview participants, then show everyone else a recording. The choice depends on whether the reviewer needs to take the turn or just see what happened when somebody else did. Google’s developer launch

Key takeaways
  1. 01
    Budget the seconds Google generates.Standard pricing is $17.50 per million streamed video output tokens. At 24 frames per second, the effective rate is $0.1008 per second with stereo audio included. Inputs are separate.
  2. 02
    Build a session with a definite end.Create an interaction on the backend, connect the viewer, set up the world and send controls. Tell the service to complete the interaction when the participant finishes.
  3. 03
    Test steering from the visitor’s location.Genie 4 serves from the US, Western Europe and Japan. A short movement and a written change to the scene can also have different response times on the same connection.
  4. 04
    Record the experience when the decision is over.Worlds can change when the camera turns away. A recording preserves the particular door, sound and visitor choice discussed in the interview and can be replayed without new model generation.

01 — Creative briefGive the visitor a real choice

In the courtyard study, put a participant at the fountain with an archway to the left and another to the right. Ask them to find the source of a sound. They can turn the camera, approach either arch and change direction. The producer can observe the first turn and ask what prompted it. A fixed camera path would answer whether the scene looks convincing. It could never tell the producer which way that participant would choose to walk.

Write the decision into the brief before spending anything. If a curator simply wants to approve the appearance of the courtyard at dusk, supply a still or a short clip. If they want five people to discover their own route, pay for five live interactions. An interview that begins with eight minutes of discussion can display a still image during that discussion and create the world only once the participant is ready to move.

Project Genie and the developer API answer different access questions. A subscriber can explore Google’s own interface. An application developer uses a paid API project to own the visitor interface, the instructions and the session lifecycle. Keep the two budgets on separate lines, especially when someone on the creative team already has a subscription. The Project Genie subscriber guide covers the consumer product’s plans and duration guidance.

02 — Output meterFrom one frame to a two-minute invoice estimate

Google prices the stream by emitted tokens. A frame accounts for 240 video output tokens. Multiply by 24 frames per second to get 5,760 tokens per second. At $17.50 per million tokens, one second costs $0.1008. Multiplying duration by ten cents is close enough to decide whether to run a meeting. Use the token rate for a spreadsheet, especially if the prototype will serve hundreds of people.

Synchronized 48 kHz stereo audio comes with the video tokens. Google lists input text and reference images at $1.50 per million tokens and text output at $9.00 per million. The table covers the video line, in US dollars under Standard pricing, and assumes a continuous 24 fps stream. Hosting the prototype, storing a recording, paying participants and delivering a saved clip are separate budget decisions.

Calculated from Google’s Genie 4 Standard pricing, November 21, 2026. Dollar figures are rounded at the end. Prompt and other non-video charges are excluded.
Stream lengthFrames at 24 fpsVideo tokensOutput (USD)
30 seconds720172,800$3.02
60 seconds1,440345,600$6.05
120 seconds2,880691,200$12.10

For the museum’s five interviews, half a minute each yields 150 generated seconds. Multiply by $0.1008 to budget $15.12 in stream output. If a producer also rehearses twice for half a minute, the seven runs yield $21.17. Using the rounded $3.02 for each interview would introduce a small rounding error. Keep the underlying rate in the sheet and round the whole group once.

Repeated views of a saved recording involve ordinary player and delivery costs. Ten people watching the same recorded half minute do not ask Genie to generate ten worlds. Ten people each pressing Start on an interactive version do. That difference determines whether publishing the prototype to a wider audience is a routine video launch or an invitation to incur new inference charges on every visit.

03 — ImplementationCreate one interaction for one participant

On the backend, create an interaction with model ID genie-4. Keep the long-lived paid API key there. For a browser viewer, mint a token scoped to that single interaction and provide the token and interaction ID to the client. The browser can then connect with WebRTC. Google also supports WebSocket for applications such as a backend that wants to process individual frames. WebRTC lets a browser play the incoming media tracks while sending control events alongside them. Google’s streaming and authentication guide

Send one world_setup event with a text description, a single landscape reference image, or both. For the museum, describe a viewpoint at the fountain, the two arches and where the first sound should come from. Establish a stable starting camera. Then send native movement and camera actions as user_action events. When the producer wants the lighting to change, send a separate text input while the world is running. Ordinary movement and a written direction for the scene are different kinds of input.

Keep the controls outside the generated picture. A real Start button can focus the viewer, show the keyboard controls and begin the timer. A real Finish button can end the session even if the model redraws the courtyard. Use an ordinary interface label for the participant number or elapsed time. Signs and clocks inside the scene can deform as the viewpoint moves. The application already knows how to display both consistently.

First frame

Google’s guide estimates roughly five to nine seconds for the first frame. Let the participant read the instructions before creating the world and show a connecting state while it starts. Price emitted video frames separately from the time the participant spends on the page.

04 — Session endWhat happens after the last visible frame?

A visitor closes a tab and the screen disappears. Google’s guide says this detaches the connection while the server can continue generating. The same problem occurs when application code merely leaves the local streaming block. While the service emits video tokens, they count. The documented way to finish is to await session.complete() so the service terminates generation and billing. Stopping playback or pausing a local timer cannot communicate that instruction.

Give every deliberate exit a path to that same operation. Pressing Finish, reaching the application’s thirty-second limit and leaving for the next interview should all request completion. Prevent a double click on Start from creating two interactions for the same visitor. Keep the interaction ID and outcome with the interview record so the team can distinguish a participant who finished from one whose network dropped. An abrupt browser close may cut off a final request before it reaches the service.

Test the awkward cases before opening the room. End one trial with the button, end another by navigating away, and disconnect the browser during a third. Compare the service’s final interaction state and usage to the participant’s observed time. If your architecture cannot verify completion after a disconnect, allow for the full session in the budget for that case. A client-side log entry that says “closed” only proves the application saw its own connection go away.

At the full frame rate, the extra ninety seconds between the intended thirty-second stop and the two-minute service cap would cost about $9.07 in additional output. The actual additional cost depends on how many frames the service emits after the viewer leaves. A few missed completions can cost more than all the planned interviews, so show successful session endings alongside session starts in the pilot review.

05 — Audience locationTry the controls from the interview room

Google lists serving locations in the United States, Western Europe and Japan. Its guide identifies North America, Western and Central Europe, and East Asia as the areas where round trips are generally most responsive. A producer operating from near a serving region can see a different experience from a participant on hotel Wi-Fi elsewhere. A relay close to a serving region can help one part of the path. The participant still has to reach it.

Ask a remote participant to press once, wait for the reaction, then turn toward the fountain. Google describes roughly 0.2 to 0.5 seconds for native movements and around one to two seconds for a text change to become visible. These are the vendor’s approximate response figures. Network conditions and scene complexity can change the experience, and a network round trip measures only one part of what the participant sees. Google’s controls guide

A slow change of light can be entirely acceptable in a museum preview. A visitor who repeatedly presses a direction because they cannot see their last movement may pass the arch before they notice it. That would compromise an interview about wayfinding. Record the device, location and input-to-visible delay during rehearsal. Include an ordinary video option if a participant’s connection makes the controls unusable, and identify those sessions separately when interpreting the study.

06 — Scene limitsA memorable visit can still be hard to reproduce

The output is landscape 1280 by 720 at a fixed 24 fps. Google documents a 120-second session limit and says object persistence is strongest earlier in the interaction. The system generates the scene over a rolling period. An object can shift after it has been out of view, and exact counts and written words can drift. Extending an interview toward two minutes can therefore buy more frames just when the scene grows less dependable. Genie 4 model specifications

That matters for the museum example. A participant who turns around to look for a third exit may find a courtyard different from the one another participant saw. Compare their reactions as exploratory observations. For an experiment that requires identical door positions, measured sightlines or an exact accessible route, use a controlled map or a conventional scene with fixed geometry. A generated tour can illustrate an idea without establishing whether an actual building meets the brief.

Live exploration
Choose where the camera goes

An interview needs the visitor to choose a doorway or look for a source of sound. Record the participant’s observation and the version of the scene they saw.

Use a short timed session.
Fixed review
Approve the design seen in a meeting

A director wants colleagues to judge the same lighting change and sound cue later. Export or capture one approved playback and distribute that recording.

Use a reusable clip.
Physical validation
Check whether a route is accessible

Door widths, positions and travel distances must be measurable and repeatable. Get them from the plans or a validated spatial model.

Use surveyed geometry.

07 — Pilot capacityReserve a place before the guest presses Start

A visitor who reaches the room and sees an error has used a research slot even if the model bill is small. Google lists concurrent active session limits of two, five and ten at paid tiers one, two and three respectively. The corresponding daily request limits are 50, 250 and 1,000. The associated project’s active limits in AI Studio matter for the actual event. In-session keystrokes and steering messages do not each consume another daily session request. Google’s Genie 4 rate limits

Consider five interviewees sitting down together on a tier-one project. Five sequential half-minute sessions fit the daily request allowance and the example video budget. They cannot all start together on the documented two-session concurrency limit. Stagger the interviews or show the recorded tour to waiting visitors until a session finishes. A link that anyone can forward needs a separate limit on how many people your application admits to paid sessions.

The Genie research preview describes what DeepMind originally demonstrated. The creative video integration guide covers the handoff when the deliverable is a finished asset. For an interactive prototype brief, ask the producer for the number of people, the longest useful visit, their locations and the choice they should be able to make. Those four details turn an open-ended world into a session you can schedule and price.

Decision

Buy the seconds in which a visitor can make a choice

For a five-person pilot, budget the rehearsals as well as the interviews, test the controls from the participant’s location and verify that Finish completes the interaction. Share a recording with the next reviewer unless their own turn at the controls will answer a new question.

Digital Applied

Scope a live interactive prototype

Build the session budget around what participants need to do and how the team will review the result.

Session designUsage controlsRegional testing
Project brief

Define one visit

  • →One participant choice
  • →One spending limit
  • →One observable finish
Answers

Frequently asked questions

At a full 24 fps, the generated video uses 172,800 output tokens. Standard pricing puts that at $3.024, or about $3.02, with stereo audio included. Input tokens and other application costs are separate.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Genie 4 in Project Genie: Who Gets Access and What Works

Google is rolling out Genie 4 to eligible AI Ultra users. See which plan qualifies, how long worlds last, when Genie 3 fits better and what you can share.

November 19, 2026 · 8 minRead
AI Development

Genie 4: What Interactive Worlds Can Actually Do Now

Genie 4 adds native audio and live control to generated worlds. Understand preview access, creative uses, simulation limits and what to test before buying.

October 27, 2026 · 9 minRead
AI Development

Nano Banana 2 Lite & Gemini Omni Flash: Image to Video

Google shipped Nano Banana 2 Lite and Gemini Omni Flash on June 30, 2026. Fast image generation at $0.034 and conversational video at $0.10 a second.

June 30, 2026 · 8 minRead
AI Development

Gemini 4 Neon: Price, Benchmarks and When to Use It

Gemini 4 Neon costs $3 per million input tokens and $15 for output. Compare four thinking levels, coding benchmarks, paid access and a sample review budget.

November 19, 2026 · 8 minRead
AI Development

Seedream 5.5: Better Edits, Transparent Assets, Real Limits

Explore Seedream 5.5’s editing and transparency claims, the production work they could remove, and the access and quality questions creative teams must settle.

November 9, 2026 · 10 minRead
AI Development

Deleting AI Agent Memory: Where Stored Copies Survive

Deleting AI agent memory takes more than clearing a chat. Map stored copies, retrieval indexes and backups, then verify what your system can still recover.

September 4, 2026 · 7 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source