Tavus previews Griffin for real-time video interaction

Tavus describes Griffin as a Human Interaction Model that processes video, audio, and speech together. The company says it returns 720p video in 320-millisecond chunks and reports that 26 of 54 people in a one-minute test believed they were speaking with a human. Griffin-Lite is a preview for trusted testers, not a general release.

For builders, this points to a different latency target from turn-based chat. An assistant must manage interruption, timing, visual context, and the cost of being wrong while a person watches. The claims need independent testing.

Prototype one short interaction with clear disclosure and an easy handoff. Measure interruption recovery, delay, user understanding, and failure cost. Wait for wider access and third-party benchmarks before using human-likeness as a product metric.