DeepMind announced Genie 4 on October 27, 2026, and within a day the phrase "world model" had been stretched in every direction. Reading the announcement itself, the interesting shift is not image quality. It is that Genie 4 is natively multimodal and agent-driven: Gemini Omni grounds the visuals, a Gemini 3.8 Flash loop observes, plans and replans in real time, and audio plus dialogue come out of the same pass as the picture.
What that changes in practice:
- You steer a running scene instead of regenerating a clip. Weather, camera position, objects and characters can all be changed mid-session.
- Physics and human behaviour are simulated rather than styled. The marble-bust-on-watermelon clip is the one worth watching, because contact physics is where generated video usually falls apart.
- SynthID watermarking is on every generated frame and audio track, which matters if a clip is going anywhere near production.
- Access is staged: trusted testers first, then Google AI Ultra subscribers via Project Genie, plus Gemini API and AI Studio partners.
What DeepMind has not published: frame rate, output resolution, model size and pricing. Any exact number you see quoted elsewhere is guesswork, which is a good reminder to check the source before repeating it.
The comparison that actually helps is Genie 4 vs Genie 3, run on the same dressing-room scene. One year apart, same prompt, visibly different handling of consistency and audio.
If you want the side-by-side table, the original demo clips and a walkthrough of how access is staged, I keep a longer, source-linked breakdown at genie4.
Top comments (0)