October 2, 2026 — San Francisco.
Tavus has unveiled Griffin, what it calls the first “Human Interaction Model” — a real-time video-to-video AI that listens, watches, talks, gestures, and reacts in one continuous stream, like a person on a video call. In the company’s own one-minute video-call study, 48% of participants believed they were talking to a real human.
Why it matters: We may be approaching the point where you can no longer trust that the person on your video call is a person. Griffin processes speech, facial expressions, tone of voice, gestures, and pauses simultaneously — full duplex, in real time. For everything from tutoring to customer support, this is a genuine capability leap. It also raises hard questions about disclosure and deepfake risk, which Tavus itself seems to recognize.
What is Griffin, exactly?
Griffin is a new class of model, Tavus says: a unified, full-duplex video-to-video system rather than a pipeline of separate avatar, timing, and speech components. Co-founder and CEO Hassaan Raza frames the idea as “human computing” — computers that adapt to face-to-face human communication instead of forcing people to learn machine commands. The company started in 2020 building personalized AI videos for sales and marketing before expanding into live video conversations, and it has raised about $64 million.
The 48% study — and the caveats
The headline number comes from a Tavus-run study: 54 people recruited through a research platform were told they’d have a one-minute video call with another participant about what they were looking forward to this year. Their “partner” was actually Griffin-Lite, a smaller variant. Afterward, 26 of the 54 said their partner was a real person — 48%. By comparison, only 1 of 41 people (2.4%) took Tavus’s previous Phoenix-4.5 system for a person in the same protocol.
The caveats are real: the study was run by Tavus itself, with only 54 participants, and the recruiting platform was unnamed. Still, the gap between 2.4% and 48% is hard to dismiss. Separately, on NVIDIA’s VideoFDB benchmark for full-duplex audio-visual conversation — which Tavus says NVIDIA ran independently — Griffin scored 3.83 out of 5, versus 3.92 for actual humans and 2.80 for the previous best AI.
Why it matters
From a “everything AI, tested” perspective, this is the first credible claim of a video Turing test pass, and the independent NVIDIA benchmark score gives it weight beyond the company-run study. Practical uses Tavus lists include tutoring, practicing difficult conversations, and camera-based tech support — applications where reading facial expressions and reacting in real time genuinely changes the experience. But the same capability is tailor-made for impersonation scams, and Tavus is notably not letting anyone use it yet: Griffin-Lite is available only to select testers as a research preview, and the company says it is withholding customer access until it ships disclosure and safety features. That restraint is the responsible move — and arguably the most interesting part of the story.
FAQ
What is a “Human Interaction Model”?
Tavus’s term for a unified, full-duplex video-to-video model that listens, watches, talks, gestures, and reacts simultaneously — instead of chaining separate models for the face, the timing, and the speech.
Can I try Griffin right now?
No. Griffin-Lite is a research preview for select testers only, and Tavus has published no pricing. The company says the more capable version won’t ship to customers until disclosure and safety features are ready.
How close is Griffin to human-level on the NVIDIA benchmark?
Very. On NVIDIA’s VideoFDB perception track, Griffin scored 3.83 out of 5 versus 3.92 for humans — closer than any previous AI, which topped out at 2.80.
Should I worry about video-call impersonation?
It’s a legitimate concern. Tavus’s own study shows one-minute calls can fool nearly half of people, which is exactly why the company is holding Griffin back from general release until disclosure safeguards exist.
Sources: The Decoder, Tech-Ish, Cellcog (via Tavus research materials)