Tavus presents Griffin as a model for real-time video interaction

Tavus says Griffin is designed to understand and generate face-to-face human interaction in real time.[2] The company describes it as combining model capabilities in one unified video-to-video system.[2] The Decoder reports that Griffin handles speech, facial expressions, tone of voice, gestures and pauses while receiving and generating video.[3] Tavus calls Griffin the first model to pass the video Turing test, a claim that should be read as the company’s characterization.[1]

Tavus says 48% of participants believed Griffin was a real person after a one-minute video call.[2] Protos reports that 28 participants identified it as AI, and that most participants who did so suspected it within 20 seconds.[4] Tavus says previous systems had a pass rate below 3%; The Decoder reports that previous systems topped out at 2% in Tavus’s comparison.[1][3] The Decoder also reports a score of 3.83 for Griffin on a test Tavus describes as an independent Nvidia evaluation, compared with 3.92 for humans and 2.80 for the previous best AI model.[3] Protos says the results were not independently verified.[4]

Protos reports that participants were recruited through an independent research platform, but Tavus has not disclosed their demographics.[4] That omission makes it difficult for readers to assess how broadly the reported result might apply across different groups. The reported 48% figure describes participants’ judgments after a one-minute call; it does not establish how Griffin would perform in longer interactions or in real-world deployments.[2][4] Journalist Kevin Roose raised a potential misuse concern, arguing that a system like Griffin could be used to trick older people into revealing bank passwords.[4] That is a criticism about possible risk, not evidence that such misuse has occurred.

The Decoder reports that Tavus is offering Griffin-Lite as a research preview to select testers.[3] Tavus plans to release a more capable version after safety concerns are addressed, according to The Decoder.[3] Tavus lists tutoring, practicing difficult conversations and camera-based tech support as potential use cases.[3] For teams assessing the technology, the distinction between a limited preview and a later, more capable release matters: the available information does not establish broad availability or confirm performance in those applications.[3] A measured evaluation should therefore focus on the team’s intended task and the evidence available for that use.