Google Genie 4 Guide: Specs, Benchmarks, Videos and How to Get Access

Google DeepMind announced Genie 4 on October 27, 2026, with a preview for trusted testers and plans to bring it to Google AI Ultra subscribers through Project Genie and to early API and Google AI Studio partners. Those wider routes are coming soon; an Ultra subscription alone does not establish access today. Google has not announced a general rollout date in the launch materials. Google announcement

For interactive prototyping and AI-agent research, the difficult question is consistency: does an environment respond sensibly when someone acts on it, and do the consequences persist? A convincing clip cannot answer that on its own.

The public evidence does not answer that question with a benchmark score yet. Google has shown selected demonstrations and described limitations, but the materials reviewed for this guide do not disclose numerical Genie 4 benchmark results, a frame rate, resolution, parameter count or API price. Kingy.ai has not tested Genie 4 hands-on.

What Genie 4 is and how it works

Genie 4 is a world-simulation model. Project Genie is a Google product through which people can explore generated worlds. The names are related, but a product’s session limit, interface or subscription terms should not automatically be treated as a specification for every version of the underlying model.

A conventional video generator produces a sequence to watch. An interactive world model must also respond to an input during the sequence. If a viewer turns left, moves toward a door or asks for a change, the next visible state needs to reflect both that action and what has already happened. A plausible response to one action is useful; preserving the consequences across a longer sequence is the harder test.

A conventional game engine normally starts from explicitly authored objects, rules and assets. Genie generates the experience, so a beautiful scene does not by itself establish that there is a reusable mesh, a persistent object database or a deterministic physics simulation underneath it. Google has not documented a Genie 4 game-engine export or a public save-state interface in the materials reviewed. Treat the output as a generated interactive experience unless the eventual product documentation establishes something more specific.

Google describes keyboard-and-mouse controls, text steering and a few minutes of interaction, with details and objects liable to drift. This is a product-page description, not a hard cap or uptime guarantee. Google DeepMind model page

Audio and scene changes

Google says Genie 4 generates ambient sound, music and responsive dialogue alongside its environments, with text edits available during exploration. Google announcement

Those features give a creator more to examine than the image alone. When a door closes, does the sound occur at the right moment? If a character moves away, does the voice appear to move with them? When a user changes the setting, does the sound suit the new setting without disrupting the action? The answers require interactive tests across many sessions, not just a strong soundtrack in one recording. Google has not established a microphone-input interface in the materials reviewed here.

Where Gemini fits

Google assigns grounding to Gemini Omni. Gemini 3.8 Flash observes frames, plans steps and checks progress. Separate Gemini benchmarks do not evaluate the complete simulation; Google has not supplied a full architecture. Google DeepMind launch

A separate example pairs Gemini 4 Argon with Genie 4 to simulate Apptronik’s Apollo playing checkers, passing steps as text captions. This does not establish physical-robot performance or sim-to-real transfer.

Google claims improved realism in physics and human movement, including aerodynamics and collisions, but supplies demonstrations rather than numerical accuracy results. Google DeepMind launch

Genie 4 specifications: what Google has and has not disclosed

The following table reflects the launch and product materials reviewed on October 27, 2026. “Not disclosed” means we did not find a value in those materials, not that Google has no internal measurement.

“Real time” describes the intended interaction; it is not a substitute for measured input-to-response latency. Similarly, a Project Genie subscription price is not a Genie 4 API price. Developers need a documented endpoint, quota, pricing and terms before they can plan a production integration.

Watch Google’s Genie 4 examples

Featured artwork: Google DeepMind, from the official Genie 4 launch.

Google’s launch film is a starting point for seeing the range of scenarios it chose to show.

Google DeepMind’s Genie 4 launch film. Watch on YouTube or see the official launch page.

These are selected Google demonstrations. Its disclaimer says the results are illustrative and sequences shortened. They do not disclose attempt counts, failure rates or behavior beyond the recording. Google DeepMind model page

Cloth: material motion

Google DeepMind demonstration: an interactive cloth scene. Open the official video.

Watch how the cloth changes shape as it moves. The useful follow-up is whether the same material continues to behave coherently after repeated actions and changes in viewpoint. A convincing fold is visual evidence, not a measurement of the fabric’s mass, friction or force response.

Train station: people and continuity

Google DeepMind demonstration: an interactive train-station environment. Open the official video.

Look at the relationship between moving people and the surrounding architecture. If a person passes behind an obstacle, does their position still make sense when they reappear? Does a landmark remain where the viewer left it? A station is a useful test because the scene contains many details that can drift independently.

Robot lab: planned action

Google DeepMind demonstration: a robotics scene. Open the official video.

Watch for the relationship between an apparent instruction, the robot’s movement and the visible result. The next research question is whether each prerequisite can be checked and repeated, including failures. The clip alone does not report task-success rates or transfer to a physical robot.

How Genie 4 compares with earlier Genie models

The family has moved from research on interactive environments toward richer and more accessible simulations, but the disclosed numbers do not form a clean head-to-head comparison. The conditions and published details differ by generation.

For more on the earlier release, see our Genie 3 guide. The figures above describe what was reported, not a measured ranking of the generations on the same tasks.

Genie 4 benchmarks: what is available?

The reviewed launch and product materials contain no numerical Genie 4 benchmarks. The inspected model-card index has no Genie entry. Google reports internal safety evaluations and SynthID watermarking for generated video and audio, without numerical safety results here. Google DeepMind launch

That is a reporting gap, not proof that no evaluations took place or that the model failed a test. It also means there is no defensible public Genie 4 leaderboard position to quote from the sources we reviewed.

World-model evaluation is an active research problem. The WorldScore paper, published at ICCV 2025, evaluates world generation across controllability, quality and dynamics, including camera-trajectory controls. Its dataset contains 3,000 examples. It illustrates why an evaluation needs to look beyond a still image. We found no evidence in that paper that Genie 4 was tested or ranked on it.

A proposed independent test protocol

The following is Kingy.ai’s proposed protocol, not a report of completed tests or scores. It is designed for a future evaluation when access and recording tools allow it.

1. Publish the setup first. Record the product and model version, region, device, connection, session limits, control method and test date. Specify the prompts, actions and scoring rules before viewing results. If a fixed seed or initial state is unavailable, say so; a retry from a visually similar scene is not necessarily an identical starting condition.
2. Sample sessions, including failures. Run a declared number of attempts across several scene types and report every attempt, including failures to start, resets and unusable outputs. Separate prompted scene changes from ordinary navigation. Publish the denominator for every success rate and retain unedited recordings where permitted.
3. Check object permanence. Place or identify a distinctive object, look away and return after fixed intervals. Score identity, location and condition separately. Repeat after a scene edit. A changed object may be intentional if the prompt requested the change, so evaluators should distinguish expected edits from drift.
4. Measure control and latency. Issue a predefined sequence of movements and score whether the observed direction and magnitude match the intended action. Measure time from input to visible response at the median and tail of the distribution, across connection conditions. Record lost inputs separately from delayed ones. Without input timestamps, do not call video playback speed a latency measurement.
5. Test cause and effect. Give a task with explicit prerequisites, such as locating an item before placing it on a marked surface. Check the state after each step. Count a visually plausible final scene as a failure if the required intermediate action was skipped. Have reviewers assess recordings against the same checklist.
6. Test audio grounding. Ask reviewers to compare sound onset, source location and timing with visible events. Change one condition at a time, such as moving away from a sound source. Keep audio-quality judgments separate from whether sound corresponds to the scene.
7. Track degradation over time. Score the same attributes at fixed intervals throughout each available session. Record resets and the longest uninterrupted run, but do not infer a service guarantee from the best run. Compare early and late observations across the full sample, not only the most impressive clip.
8. Separate visual plausibility from physical accuracy. Where a task makes numerical physical claims, compare observable trajectories or contact behavior with measured reference data under stated assumptions. A scene can look plausible while violating the forces or geometry needed for engineering work.

For meaningful comparisons, another system should receive the same tasks, scoring rubric and reporting rules. Report uncertainty and failed attempts alongside averages. If the two systems expose different controls or session lengths, document those differences instead of presenting the result as a universal model ranking.

How to get access to Genie 4

The preview and planned subscriber/partner routes are described at the top of this guide. The next step for interested readers is the application below; admission and timing remain Google’s decision.

The practical route for a prospective tester is to apply using the official Genie 4 access form, then check the version and access status in Project Genie when invited. An application is not a guarantee of access. Do not buy an Ultra subscription on the assumption that it grants immediate Genie 4 access.

Google’s existing Project Genie help describes a Genie 3 experience, including a 60-second exploration session. Those instructions are useful for the existing product, but they do not establish Genie 4’s interface, quota or session limit. Google previously expanded Project Genie to eligible adult Ultra subscribers globally; eligibility and local pricing should be checked in the product rather than inferred from that earlier announcement. Project Genie expansion

For developers, the next useful artifact is official partner documentation: authentication, an actual model identifier, inputs and outputs, latency, pricing, quotas, data terms and error behavior. Until those exist publicly, an SDK snippet with a guessed model name would be speculation.

Why this matters for creators and researchers

For a creator, the immediate appeal is iteration. A team exploring a location, camera path or mood could potentially navigate an idea and make changes while watching the result. That might help with storyboarding, previsualization or early concept discussion. The practical value depends on how reliably the system preserves a chosen scene, how long a session lasts and whether the output can be used under the product’s terms. A generated walkthrough should not be presented as a finished game or a set of exportable production assets without documentation establishing those capabilities.

The related Google Playground guide covers a different product and workflow. A game-creation tool can produce editable game elements; Genie 4’s public materials describe a generated interactive simulation. Check the actual output and editing options before choosing either for a project.

For agent research, a world model could provide varied environments in which a planner observes a scene, chooses actions and receives feedback. Useful experiments would measure whether those actions lead to the intended state and whether changes persist. Researchers also need to know what the system hides: if an evaluator cannot inspect the state or reset the environment, it becomes harder to distinguish a planning error from a simulation inconsistency.

Robotics has a higher bar. Training or evaluating a robot in a generated scene requires evidence about geometry, contact, timing and the correspondence between simulated and real actions. A credible transfer study would compare a policy trained with the generated environment against appropriate baselines on held-out physical tasks, reporting success and failure rates. Google’s simulated robot demonstration makes the research direction visible; it does not provide that transfer evidence.

Current limitations and the right first experiment

Google acknowledges limitations involving persistent details, locations, interactions among multiple agents, text and audio. Google DeepMind model page Those limits matter differently by use. A slightly inconsistent background may be acceptable in a mood board; an invented object location can invalidate an agent’s task. A fluent spoken line can still be mistimed or disconnected from the person on screen.

If you receive access, start with one small task whose result you can inspect: choose an environment, identify a landmark, move away and return to it. Record the attempt, repeat it, and note any changes before adding a text edit or a multi-step goal. That gives you a clearer baseline than judging the system by its most polished scene.

The Kingy Brief

Get The Kingy Brief.

AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.

Free · Double opt-in · Unsubscribe anytime

Privacy policy

Sign up for The Kingy Brief on Beehiiv.

Signup help and newsletter schedule

Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.

Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.

Sign up free