Categories / Video & Audio
Griffin-Lite · Research previewTavus Griffin: Real-Time Video Interaction
Tavus Griffin is a real-time, full-duplex video-to-video interaction model. Griffin-Lite is a research preview for selected trusted testers, not a generally available customer product.
Official-source review · Last checked October 2, 2026 · No independent product benchmark
A live conversation, not a rendered talking head
Tavus describes a system that keeps watching and listening while speaking, responds to visual cues and handles interruptions. Its continuous perception and audiovisual response distinguish the intended task from prerecorded avatars or standalone text-to-speech.
Research preview access
Griffin-Lite is not available for customer use at this time. Tavus offers a request route for selected trusted testers. General release timing, commercial pricing and customer API access are not established by this preview. Check Tavus access and safety notes.
What the 48% result means
Tavus reports that 26 of 54 participants (48%) thought their Griffin-Lite partner was human after a one-minute video call. Tavus ran/reported the study; recruitment used an independent research platform. The same protocol with Phoenix-4.5 + Sparrow-2 + Raven-1 involved 41 people; one (2.4%) answered human. Participants were informed afterward.
AI Tool Finder did not independently reproduce this study. It is not evidence that every conversation will appear human, nor a general quality ranking. This face-to-face study is separate from NVIDIA’s VideoFDB evaluation; their scores are not interchangeable. Study and methodology.
Before evaluating a live-video agent
- Use it only if continuous audiovisual interaction is central to the task; a recorded explainer does not require this model.
- Plan clear AI disclosure, camera/microphone consent and a way to end a session.
- Test interruptions, silence, visual ambiguity and recovery in your actual environment. A polished demo does not measure your network or session reliability.