Tavus's Griffin-Lite fools 48% in video calls

Tavus built Griffin as a so-called Human Interaction Model. Unlike typical voice AI that waits for a user to finish before recognizing, generating a reply, and synthesizing speech, Griffin continuously reads video and audio and can nod, murmur, or interrupt in under a second. It generates not just a face but arms, fingers, body motion, chair, shadow, and background from one reference image, at 720p in 320-millisecond chunks, with an average 0.43-second delay from received audio to on-screen video on NVIDIA H100 hardware. Tavus says that is about half the delay of the next-fastest streaming video generation method it compared.

On NVIDIA's VideoFDB benchmark, Griffin-Lite scored 3.83 out of 5 on Generation, close to the 3.92 for human reference data and above Gemini 2.5 + Anam at 2.80. On Perception it scored 3.73 versus 4.20 for humans and 3.44 for MiniCPM-o 4.5; Tavus says NVIDIA ran that evaluation independently in September 2026. In the conversation study, participants rated Griffin-Lite 5.4 for naturalness, 5.6 for trustworthiness, 5.8 for wanting to talk again, and 5.5 for feeling listened to, while natural conversational flow scored lowest at 4.9.

The gap between the 4.9 flow score and the 48% human judgment is the tension Tavus is sitting on. Whether Griffin moves beyond a limited research preview is likely to hinge on how well the company can build visible signs that a caller is an AI without degrading the very naturalness that produced these results.