Compare the announcements on their own terms
Genie 4 and Genie 3 both belong to Google DeepMind’s interactive world-model research. The Genie 3 announcement appeared on August 5, 2025; Genie 4 followed on October 27, 2026.
The useful comparison is what each publication actually documents. Neither a release number nor a promotional recording tells you how often a system succeeds on your own task. We have not run these models in a controlled comparison.
A source-backed capability comparison
| Dimension | Genie 3 announcement | Genie 4 announcement |
|---|---|---|
| Interaction | Real-time exploration of generated worlds | Real-time simulation with natural-language steering |
| Audio | Does not establish the new native-audio capability | Introduces native ambient sound and interactive dialogue |
| People and physics | Published demonstrations of interactive environments | Highlights more natural human action and physical realism |
| Reasoning | Do not infer the later orchestration design | Describes Gemini-assisted planning and scene steering |
| Numeric display specification | 24 fps and 720p stated at announcement | Do not transfer those figures without a Genie 4 specification |
| Continuity | A few minutes of consistency described | Longer-term consistency remains a research limitation |
| Access | Historical research access at its launch | Trusted-tester preview; broader routes described as upcoming |
A blank or qualified entry means the cited material does not establish the claim. It does not prove that a capability is absent.
Native audio changes what needs to be evaluated
A silent scene can be judged primarily by its visual continuity. With native audio, the evaluation also needs to consider whether sounds match the visible setting and respond coherently to changes.
For example, a door closing creates a clear test event: note its visible closure and the associated sound. Then move away from the door and check whether the environment remains acoustically plausible. This is an original evaluation suggestion, not a reported measurement of either model. The audio guide provides a listening worksheet.
Steering a world is different from authoring deterministic rules
Real-time edits let a user redirect a generated experience. They do not, by themselves, provide the same contract as editing a game’s scene graph, collision system or save data.
Use a repeatable route and one change at a time when comparing behaviors. If one session introduces three people, rain and a moving camera while another uses an empty room, their apparent quality cannot be attributed solely to the model generation. Match the scene brief, actions and observation window as closely as practical.
Do not reuse old limits as new specifications
A common comparison error is to copy Genie 3’s frame rate, image resolution or interaction duration into a Genie 4 table. A number needs a version, a source and a scope. A prototype’s session cap is also different from a model’s underlying continuity.
The same caution applies to account access. A tutorial showing Project Genie in early 2026 may be accurate for that experience at that time without describing the new research preview. Use the current access guide before deciding what to subscribe to or test.
Choose a comparison question you can answer
Start with a narrow question: does a landmark retain its appearance after a turn, does an instruction preserve the viewpoint, or does a sound follow an observable event? Save the prompt, actions and recording together.
Report the observed cases and their limitations. “This object changed during this run” is more informative than an unsupported reliability percentage. If you need repeatable mechanics, editable assets or production export, check a separate toolchain contract; a world-model comparison does not answer those requirements.