Genie 3 (left) vs Genie 4 (right)
Both models received the same scene. Look at how the characters resolve, how the wardrobe reads, and whether the room keeps its layout as the camera moves.
Source: Google DeepMind, Genie 4 announcement
Genie 4 vs Genie 3
One year apart: Genie 3 arrived in August 2025 with real-time interaction at 720p and 24 fps. Genie 4 arrived in October 2026 as a multimodal, agent-driven simulation.
Here is exactly what changed, and what did not.
The short version
Genie 3 proved that a generated world could be interactive. Genie 4 adds three things on top: sound, a reasoning agent that keeps the scene coherent across interventions, and actual routes to access beyond a research preview.
- New: native audio — Ambience, music-like sound, and reactive character dialogue.
- New: agent loop — Gemini 3.8 Flash observes, plans, verifies, and replans inside the scene.
- New: access paths — AI Ultra via Project Genie, plus Gemini API and AI Studio partners.
- Same: core interaction idea — Both let you steer a generated world in real time rather than watch a clip.
Genie 3's public figures — 720p, 24 fps, several minutes of consistency, one minute of visual memory — were published for Genie 3. Google has not published equivalents for Genie 4.
The comparison clip
DeepMind published a side-by-side: the same dressing-room scene generated by Genie 3 and by Genie 4. Watch what the newer model holds that the older one does not.
This is the only direct A/B footage DeepMind published. Everything else in the announcement shows Genie 4 alone.
Full comparison
Every dimension where the two models can be compared from public information.
| Dimension | Genie 4 | Genie 3 |
|---|---|---|
| Announced | October 27, 2026 | August 5, 2025 |
| What it is | Real-time, natively multimodal world model | Generative interactive environment model |
| Output modality | Visuals, ambient audio, music-like sound, live dialogue | Visual frames |
| Resolution and frame rate | Not published | 720p at 24 fps |
| How it is steered | Natural-language edits to weather, camera, objects, people | Real-time navigation plus text-triggered world events |
| Reasoning | Gemini Omni grounding; Gemini 3.8 Flash agent loop that checks frames and replans | Not described as agent-driven |
| Consistency | Longer sessions claimed, but drift still reported after many edits | Largely consistent for several minutes; visual memory about one minute |
| Character behaviour | Natural movement, reactions, and interactive dialogue | Character movement described; no dialogue generation |
| Watermark | SynthID on video and audio | SynthID on video |
| Availability | Trusted testers, AI Ultra via Project Genie, Gemini API and AI Studio partners | Limited research preview for a small cohort of academics and creators |
| Multi-agent interaction | Not addressed in the announcement | Explicitly called an unresolved research challenge |
| Agent training focus | Explicit — robotics and agent training named as targets | Explicit — training and evaluating agents including SIMA |
Blank or "not published" cells are deliberate. Filling them with guesses would make this table useless.
Scroll the table sideways to see all columns.
The three changes that matter
Beyond a version bump, Genie 4 changes what the model is for.
Genie 3 produced pictures. Genie 4 produces a place with ambience, tonal material, and characters that speak. That single change moves it from visual prototyping toward something closer to a scene you can shoot in.
What did not change
The limitations carried over. Genie 4 is better, not solved.
Drift is still real
Google still reports loss of visual detail and object permanence over long sessions or after many edits.
Physical accuracy has a ceiling
Complex interactions remain a documented weakness — the same category of problem as Genie 3.
Text on screen is unreliable
Legible writing only appears reliably when it is explicitly prompted.
Not a geography engine
Real locations cannot be reproduced with perfect spatial accuracy.
Sessions are not marathons
Neither generation is built for hours of continuous exploration.
Multi-agent play is unsolved
Genie 3 called multi-agent interaction unresolved; Genie 4 does not claim to have solved it.
So is Genie 4 a big jump?
For researchers and studios, yes: audio plus an agent loop turns a visual toy into a testable environment. For anyone expecting a public product today, not yet — access is still gated and no pricing exists. For anyone comparing specifications, the honest answer is that Google has not published Genie 4's frame rate or resolution, so a spec sheet comparison is not possible yet.
- Biggest practical gain — Coherence across interventions, because the agent verifies its own steps.
- Biggest practical gap — No published specs and no open access.
- Most overhyped claim — Treating it as a finished video tool. It is a simulation.
Comparison questions
The specifics people ask when deciding between the two.
See the capabilities in detail
The features page breaks down steering, physics, audio, and event simulation — the four areas where Genie 4 actually differs.