Tavus launches Griffin: A 1-minute video chat with 54 participants, 26 of whom mistook the other side for a real person.

Beating AI News Flash: AI video firm Tavus has launched Griffin, billed as the first Human Interaction Model focused on real-time face-to-face communication. Griffin can see and listen while speaking, reacting in real time to the other party’s expressions, movements, and pauses, and pauses immediately if a user interrupts mid-sentence. Tavus recruited 54 participants from the U.S. and Europe to hold 1-minute video calls with Griffin-Lite, without disclosing in advance that the counterpart was an AI. The result: 26 participants, or 48%, believed they had chatted with a real person. By comparison, Tavus’ prior generation model Phoenix-4.5 saw only 1 out of 41 participants (2.4%) make the same judgment error in a similar test. Tavus thus claims Griffin is the first model to pass the "real-time video Turing Test". However, the test was self-organized by Tavus with a sample size of just 54, and no larger independent replication has been conducted to date. Griffin-Lite also topped NVIDIA’s VideoFDB full-duplex audio-video evaluation, scoring 3.83/5, close to the human benchmark of 3.92. The previous highest public baseline, Gemini 2.5 + Anam, scored 2.80. Griffin-Lite is currently only available to a small group of testers and has not yet been officially offered to regular customers.