Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?
Checked 5 October 2026 · Tavus research preview
Griffin's 0.43-second face is not a 0.43-second conversation.
Provider says: Griffin-Lite listens and watches while it generates a speaking face. Tavus reports 0.43-second average audio-to-video generator latency on H100s. That times a component, not a customer call.
Two clocks, two tests
Tavus says its video generator produces 720p in 320 ms chunks: eight frames at 25 fps per latent. Its 0.43-second figure measures incoming audio to its effect on generated video in an audio-to-video comparison.
Independent benchmark operator: NVIDIA VideoFDB lists Griffin Lite at 3.83/5 for generation and 3.73/5 for perception. Its separate median event-timing rows are 1,892 ms and 2,232 ms. Do not substitute either benchmark clock for the time from joining a call to seeing an answer. VideoFDB scores 237 clips with a model judge; it is not a billed production session.
Can a builder use it now?
Not as a customer product. Tavus says Griffin-Lite is a research preview for selected trusted testers while it works on safety and disclosure. It has not published a Griffin-specific customer price. Its existing Conversational Video Interface uses other models; its listed minutes cannot price Griffin.
For a coverage decision, the useful angle is the full-duplex behavior and the three separate evidence boundaries: generator latency, benchmark conversation timing, and a real user's delivered session. For a build decision, keep Griffin on the watchlist. Ask for public access, a disclosed session, receiver-visible response and an invoice before ranking it against a route you can deploy.