Tavus ships Griffin; in its own test nearly half mistook it for human

Tavus released Griffin, a full-duplex video interaction model, on October 1. In the company's own test, 48% of participants believed they were talking to a human; its previous system scored 2.4% on the same test.

Griffin abandons the cascaded pipeline of speech recognition, language model, speech synthesis and avatar rendering. A single model handles listening, speaking, expressions and pixels at once, generating 720p video in real time with an average response latency of 0.43 seconds. On NVIDIA's VideoFDB full-duplex benchmark, Griffin-Lite scored 3.83 on generation, close to the human reference of 3.92.

The 48% figure comes from a test Tavus designed itself: participants were led to believe they were on a call with a human, the sample was 54 people over one-minute calls, and the protocol was not a standard Turing test — community notes on X flag it as independently unverified. Those who grew suspicious mostly saw through it within 20 seconds, and Griffin-Lite is open only to a small set of trusted testers, not general users.