Tavus previews Griffin for live AI video conversations

Tavus introduced Griffin on October 1. Its research preview gives selected testers a system that sees, hears and generates speech and video during a conversation.

Tavus says Griffin handles interruptions and gestures while continuing to listen and watch.

Separate the benchmark from the human study

The NVIDIA VideoFDB leaderboard lists Griffin Lite at 3.73 for perception and 3.83 for generation on its zero to five scale. Human reference scores are 4.20 and 3.92 respectively. These tracks examine understanding conversational cues and producing appropriate responses.

VideoFDB uses 237 clips from real video calls, covering eleven conversational dynamics. A language model judge applies scoring rubrics. The leaderboard compares different kinds of systems, including some that return only audio and others that generate video. Its aggregate quality scores should not be read as a universal ranking of every conversational capability.

Tavus reports that after one minute calls, 26 of 54 participants believed Griffin Lite was a person. They expected another participant and learned afterward that it was AI.

That is a narrowly defined study outcome. It does not show that the model is indistinguishable from a human across longer calls, different settings or informed participants. The conditions of the encounter matter as much as the percentage in the headline.

Customer access remains closed

Customer access remains closed while Tavus develops safety and disclosure measures. The announcement gives no firm date for general availability or price.

For teams evaluating video agents, the useful next evidence is sustained interaction testing with clear AI disclosure. A convincing face and voice can improve communication, but they do not establish factual reliability or suitability for a particular deployment.

Featured image is an original AI generated conceptual editorial illustration.