Tavus previews Griffin model for full-duplex video conversations
The AMW Read
Griffin meaningfully updates Tavus's real-time audiovisual interaction capabilities, but a screened preview and short company-run test limit demonstrated impact to the conversational-avatar subsegment.
Tavus previews Griffin model for full-duplex video conversations
Tavus unveiled Griffin on October 1, introducing a conversational model that processes live audio and video and generates responses through parallel conversation and audiovisual engines. Griffin-Lite is available only to screened research testers. In a company-run experiment, 26 of 54 participants believed their one-minute AI video call was with a human, compared with one of 41 using Tavus's previous system. Tavus calls this a real-time video Turing test, but the reported result establishes performance within that short, company-organized experiment.
The market significance sits in interactive generative media: competition extends beyond avatar appearance to perception, interruption handling, gaze and synchronized speech. Griffin reassesses conversations at sub-second intervals, allowing responses while users are still speaking. Tavus reports that Griffin-Lite led perception scoring among 15 systems evaluated on NVIDIA's VideoFDB benchmark and averaged 0.43 seconds of audio-to-video latency on H100 chips. Those results suggest progress in conversational responsiveness; they do not establish commercial adoption or reliability across longer interactions. The release therefore updates the audiovisual interaction layer without resolving whether these systems can sustain useful production conversations.
For builders and investors, the concrete diligence task is to test responsiveness alongside identity disclosure over longer sessions. Tavus says it is refining safety and disclosure features before public availability because natural interaction can cause users to mistake AI for a person. The research preview makes that tradeoff consequential: benchmark performance and human resemblance need to be assessed together with whether users consistently understand whom they are speaking with.
