Tavus Releases Griffin Model, 48% of Participants Mistakenly Believe It's a Real Person

AI startup Tavus released its new model Griffin on October 1, calling it the first "human interaction model" capable of understanding and generating face-to-face conversations, focusing on expressions and pauses. In a test, 54 participants had a one-minute video call with Griffin-Lite, and 26 of them (48%) believed the other party was a real person, while only 1 out of 41 participants believed so in the previous generation system. Participants were asked at the end of the call if they suspected the other party was not a real person, typically noticing doubts within 20 seconds. This result comes from Tavus's own research page, with participants recruited through its independent research platform. Community comments noted that these results have not been independently verified. In NVIDIA's VideoFDB benchmark, Griffin-Lite ranked first, scoring 3.83 in generation dimension, while the next best system scored 2.8, and human reference was 3.92; the perception dimension score was 3.73, with the strongest baseline at 3.44 and human reference at 4.2. Tavus stated that NVIDIA independently conducted this evaluation. Griffin supports full duplex, able to listen, see, and speak simultaneously, with an average audio-video delay of 0.43 seconds on NVIDIA's H100 chip. Tavus completed a $40 million Series B funding round led by CRV in November 2025.