Tavus says 48% of testers mistook its Griffin model for a human

Tavus says its new Griffin model is the first to pass a real-time video Turing test, after 26 of 54 participants in a live study believed they had spent a minute talking to a person rather than an AI.

The San Francisco company announced Griffin in a blog post on October 1, written by co-founder and CEO Hassaan Raza and head of research Ioannis Patras. Participants were told they would be paired with another participant for a one-minute video call about what they were looking forward to this year. Their partner was a Tavus PAL running Griffin-Lite. Asked at the end whether that partner had been a real person, 48% said yes. On Tavus's previous conversational video setup, built from Phoenix 4.5, Sparrow-2 and Raven-1, 1 of 41 participants said the same.

Griffin is a full-duplex video-to-video system, which means it watches and listens while it speaks. Tavus says it reassesses the state of the conversation at sub-second intervals rather than once per turn, so it can nod, backchannel mid-sentence, begin a reply before you finish, or stop the moment you cut in. Tavus's post describes Griffin coaching someone through a Rubik's Cube while they turn it, and holding a long pause instead of talking over it.

The measurement is the part a buyer can check. NVIDIA scored the system in September 2026 on VideoFDB, its benchmark for full-duplex audio-visual conversation. Griffin took 3.83 out of 5 on the generation track, against a human ground-truth reference of 3.92 and 2.80 for the next-highest system, Gemini 2.5 with Anam. Tavus reports a 62.8% takeover-rate alignment on that track, the highest of any system NVIDIA evaluated.

Tavus builds PALs, the Personified Application Layers its customers put in front of their own users through an API. Tavus says 150,000 developers and businesses already use the platform.

Griffin is not on it yet. Griffin-Lite is a research preview open to selected testers, and Tavus says the wider release waits on safety work, including disclosure features. A model that can pass as a person on a video call is also a model that can deceive one, and Tavus acknowledges that in the same post. For anyone building with avatars, the labelling argument that followed SynthID and Europe's transparency rules is likely to arrive here on the same schedule.

Griffin-Lite is open to selected testers through Tavus's request form; the full model has no announced date or price.