Tavus Unveils Real-Time Video Interaction Model Griffin, Says 48% of Callers Thought It Was Human

In October 2026, AI research team Tavus released Griffin, a real-time video interaction model, alongside a research preview called Griffin-Lite. In an experiment, 26 of 54 participants who had a one-minute video call with Griffin-Lite believed they were talking to a real person, roughly 48%. By comparison, only 1 of 41 participants believed they were talking to a real person with the previous system, or 2.4%. Tavus said this makes Griffin the first model to pass a real-time "video Turing test," and placed it in a new category it calls Human Interaction Model (HIM).

According to Tavus, Griffin uses a full-duplex video interaction design, meaning it can keep receiving the user's voice and video while generating its own speech and visuals. The system consists of a continuous dialogue modeling engine, which assesses the state of the conversation at intervals under one second and outputs control signals such as emotion, expression and motion, and an audio-video generation engine, which turns those signals into continuous sound and imagery. Griffin-Lite generates 720p video at 25 frames per second in 320-millisecond segments, and its speech module streams audio and can clone a voice from about 10 seconds of reference audio.

In Tavus' H100 tests, the average delay from audio arrival to the effect appearing on screen was 0.43 seconds. On Nvidia's public VideoFDB benchmark, Griffin-Lite scored 3.83 in the generation track, versus a human reference score of 3.92, and 3.73 in the perception track, versus 4.20 for humans, ahead of other AI systems listed in both tracks, though some comparison inputs differed. Tavus has opened the research preview only to selected trusted testers, not general customers; the team says it is developing AI identity disclosure features and continuing safety and alignment evaluations.