A new generation of real-time AI video models has achieved what seemed impossible: generating content faster than it can be watched. Fal’s MiniMax H3 Max model can now produce 5-15 second video clips in less time than playback duration—spawning “never-ending” generative video experiences where viewers vote on what happens next.
The breakthrough has already spawned several innovative applications: Fal Live, Levels.io’s Infinite Slop, and Reactor all offer interactive channels where AI generates continuous narrative based on audience input. One creator built a choose-your-own-path video experience using an open-source template and an AI coding assistant.
But the economics tell a cautionary tale. Because the system generates multiple possible outcomes in parallel and discards those the viewer doesn’t select, each interactive session costs roughly $15—well outside consumer pricing viability. That single data point matters more than the demo itself: it’s a concrete cost benchmark that most coverage of real-time generative video omits.
The speed jump is appearing across the AI stack. Runway’s newly announced Solaris model (limited release) generates entire interfaces on the fly rather than serving pre-built UI—potentially personalizing software per user. An open-source operating system called Omarchy (from the team behind Ruby on Rails) lets users request new features conversationally rather than through settings menus.
Perhaps most significantly, a company acquired by AMD is running language models at roughly 10,000+ tokens per second—an order of magnitude faster than typical chat interfaces. If this capability becomes standard, “instant” AI assistance transforms from marketing claim to experiential reality.
The immediate question is economic: real-time generative video is impressive but currently expensive per session. Commercial viability likely requires either dramatic cost reduction or novel business models that absorb the per-session generation costs. Monitor pricing developments, but don’t plan budget around real-time video production just yet.
What seems increasingly inevitable is that the bottleneck shifts from generation speed to creative direction. When machines can outthink human attention spans, the human role becomes curation and intent—not production.