Tavus says 48% of viewers thought Griffin was human

Tavus introduced Griffin, which it calls the first Human Interaction Model: a full-duplex video-to-video system where perception, turn taking, speech and video generation all run concurrently rather than in a request-response loop.

The headline result is a deception rate. After a one-minute live video call, 48% of participants (26 of 54) believed the partner was human, against 2.4% (1 of 41) for the previous stack of Phoenix-4.5 with Sparrow-2 and Raven-1. On NVIDIA’s VideoFDB benchmark Griffin scores 3.83 for generation against a human reference of 3.92 and a next-best of 2.80, and 3.73 for perception against a human 4.20 and MiniCPM-o 4.5 at 3.44, making it the most perceptive of 15 models tested. NVIDIA scored these in September 2026. Takeover-rate alignment is 62.8% on generation and 73.8% on perception.

Technically it runs audio-to-video latency of 0.43 seconds on H100s, about half the next fastest published streaming diffusion baseline, generating 720p in 320ms chunks at 25fps using three diffusion steps per latent with 8x temporal VAE compression. It ranks first on DOVER, FID and THEval, and second on LSE-C at 7.27.

The access situation is the important part, and Tavus is direct about it. Griffin-Lite is a research preview for a select group of trusted testers and explicitly will not be available to customers at this time. It is not on the Tavus platform. Tavus says the first model to pass the video Turing test creates a deception risk, that further alignment, safety work and safe-disclosure features are required, and that a general release follows only after those are addressed. The page also states its timing diagrams are illustrative rather than measured from real sessions, and hedges the Turing test claim with “to the best of our knowledge.”

(Original source: Tavus)