Overview
- Google announced Genie 4 as a limited research preview on Tuesday, October 27, 2026, introducing native audio, more lifelike human motion and roughly two minutes of scene consistency.
- Genie 4 uses Gemini models as in-loop directors that watch frames, break goals into steps and steer simulations in real time to keep scenes on track.
- Users can edit worlds mid-run with natural-language prompts to change camera angle, weather or add people and objects while audio and visuals update together.
- Google cautions that Genie 4 still drifts over longer runs, can lose object permanence and shows imperfect physics and readable in-world text, which limits its use for high-stakes robot validation or full game replacements.
- Access is restricted to trusted testers now with planned expansion to Google AI Ultra subscribers and select partners, and all outputs will carry SynthID watermarking under Google’s stated safety guardrails.