Tavus Unveils Griffin, a Real-Time Video Conversation Model: 48% Mistook It for a Human in One-Minute Calls

US AI research company Tavus announced Griffin, a Human Interaction Model (HIM) built for real-time two-way video conversation, on October 1, 2026. Its core is a full-duplex video-to-video architecture that listens while it talks and reacts while it listens. Tavus presents it as the first instance of a new model category, the HIM.

Image source: Tavus

The announcement rests on two numbers. In Tavus's own blind test, 26 of 54 participants (48%) in one-minute video calls believed the other party was a real person, and on NVIDIA's full-duplex AI video benchmark (VideoFDB), Griffin scored 3.83 against a 3.92 human reference, ranking first. The 48% figure comes from Tavus's own study, not from a standardized independent test. Because the model can be mistaken for a real person, general release is on hold, and only a Griffin-Lite research preview is available on a limited basis.

What the 48% video Turing test and the NVIDIA VideoFDB top rank mean

Tavus claims Griffin is the first model to pass a real-time video Turing test. The test procedure was a one-minute live blind video call with a Griffin-Lite-powered system, in which 26 of 54 participants (48%) said they had been talking to a real person. As a control, Tavus's previous-generation system under the same procedure fooled only 1 of 41 people (2.4%), a sharp gap from the sub-3% pass rate of earlier systems, according to Tavus.

On NVIDIA's VideoFDB benchmark for full-duplex AI video, Griffin scored 3.83 and topped the public leaderboard. With the human reference at 3.92, Tavus describes the score as the highest ever recorded for an AI and close to human level. Both figures originate from Tavus, and the video Turing test portion uses a company-designed experiment rather than a standard protocol, so readers should weigh that accordingly.

Beyond face lip-sync: generating the full frame in full duplex

Griffin generates the entire video frame rather than synthesizing only the face region. Tavus says it continuously senses the other party's audio and video at sub-second intervals, responding to speech, expressions, hand movement, and silence timing, so that what it sees can change what it is saying. Full duplex means both sides can talk and react at once, interrupting, nodding, and overlapping the way people do, instead of taking turns like a walkie-talkie.

The demo thread Tavus posted on X makes this concrete. In a "Simon Says" game, the model must hear the "Simon says" cue, watch the players' hands, and move its own hands at the same time. In a Rubik's cube coaching scene, it guides from what it sees in the other person's hands and waits silently when they go quiet to think. In a soldering walkthrough, it talks the worker through the job while counting seconds, and when shown a peacock figure mid-story, it writes the object straight into the plot. Emotional expression is likewise designed to surface in voice, face, and gesture at once, according to Tavus.

Public release on hold; Griffin-Lite preview only

Reaching a level where it can be mistaken for a real person is itself the reason for the release constraint. Tavus says Griffin is the first model people can mistake for a real person, so safety work must come before a public release, and the full launch is on hold. For now, only a screened group of testers has access to the Griffin-Lite research preview, and the research preview document covering how the model works, how it was tested, and which safety measures must come first is available on the official Griffin page. General customer availability stays on hold until AI-identification and misuse-prevention safeguards are in place.

Sources

• Tavus official page: Griffin: The First Human Interaction Model
• Tavus on X (@tavus): Introducing Griffin, the first model to pass the video Turing test