Tavus Says Its Griffin Model Passed a Video Turing Test, With Strong Caveats

Tavus announced Griffin on 1 October 2026, describing it as the first Human Interaction Model. It is a real-time, face-to-face video-to-video system that combines listening, conversational decisions, and expressive video and speech generation in one model. Only a limited research preview, Griffin-Lite, exists so far, and it is not on the Tavus platform.

The headline claim is a video Turing test. In Tavus's own study, participants joined a one-minute video call to talk about what they were looking forward to this year, without being told their partner was an AI. Of 54 participants, 26 (48 percent) believed they were talking to a real person. The earlier Phoenix-4.5 model scored 2.4 percent with 41 participants.

The sample is small and the study was run by the company, so treat the number as a signal rather than proof. Tavus says so itself. Participants rated naturalness at 5.4 and trustworthiness at 5.6 on a seven-point scale. Conversational flow scored lowest, at 4.9.

On speed, Tavus reports that the video generator averaged 0.43 seconds from audio input to visible effect on H100 GPUs, with a best case of 0.27 seconds and a worst case of 0.59. It says that is about half the latency of the next-fastest method it tested. It also reports ranking first on NVIDIA's VideoFDB benchmark, which is a company-reported result.

What stood out to us is the safety note. Tavus acknowledges that the qualities that make Griffin convincing could lead people to believe it is not AI, and says further safety work and safe disclosure features are needed before release. We agree, and we think disclosure should be designed in from the start, not added later.

For brand and web work, this points to a future where a conversational face on a site is realistic enough that honest labelling becomes a design requirement. We would build that label into the interface, not hide it in the terms.

Access is by waitlist for now, and pricing is not listed.