Tavus launches Griffin: video interaction model fools 48% of study participants in 1-minute calls

Tavus launched Griffin on October 1, describing it as the first Human Interaction Model (HIM) and unifying real-time perception, conversational decision-making, and expressive video generation into one end-to-end video-to-video system. Griffin is the first model to pass Tavus's real-time video Turing test: in the company's own study, 54 participants each held a 1-minute video call with Griffin-Lite believing they were speaking with another participant; afterwards, 26 (48%) believed they had spoken with a real person, versus only 2.4% for Tavus's prior systems under the same setup. Confidence was 79% among those who judged the partner human and 81% among those who judged it AI.

On NVIDIA's full-duplex audiovisual conversation benchmark VideoFDB released in May, Griffin-Lite reached 3.83 on the generation track (human reference 3.92, next-best Gemini 2.5 + Anam cascade 2.80); on the perception track it scored 3.73 (human reference 4.20, highest public baseline 3.44). On H100, Tavus measured 0.43 seconds from audio input to visible response, about half the next-fastest method. Griffin-Lite ranked first on DOVER, FID and THEval image-quality metrics and second on lip-sync LSE-C.

Griffin-Lite is not yet available to Tavus's commercial customers, which include 150,000 developers and enterprise customers such as Amazon, Salesforce, and Mayo Clinic; it is being offered only to a small set of trusted testers. Tavus is building mandatory AI-identity disclosure features and working with AI safety organizations to meet the EU AI Act Article 50 requirement, and plans to apply Griffin to one-on-one tutoring, employee conversation rehearsal, and similar scenarios.