Tavus Griffin video Turing test

Tavus launched Griffin, a full-duplex video-to-video AI it calls the first 'Human Interaction Model' (HIM), on or around October 1, 2026. In a company-run study of 54 participants, 26 (48.1%) believed Griffin was a real human during a live one-minute video call, compared to 1 of 41 (2.4%) for Tavus's prior Phoenix-4.5 + Sparrow-2 + Raven-1 stack. Griffin scored 3.83 out of 5 on NVIDIA's VideoFDB benchmark for generation, with a perception score of 3.73, below a real human's 3.92 but above the next-best published system's 2.80. Unlike earlier systems that chained separate models for hearing, thinking, speaking, and animating, Griffin handles all of these in a single unified system, processing speech, facial expressions, tone of voice, gestures, and pauses simultaneously. Tavus is withholding general customer access, stating that further alignment and safety procedures are required and that it is developing safe disclosure features while working with AI safety organizations.