According to CoinMeta, Tavus recently released its new model Griffin, which claims to have successfully made 48% of participants in video calls mistakenly believe they were talking to real humans. The model was unveiled on October 1st and is designed to understand and generate face-to-face conversations, focusing on expressions and pauses. Compared to previous systems, Griffin has seen a significant improvement in performance; the previous system only achieved a 2.4% success rate in the same test. Tavus explains that participants were informed before the call that they would be having a one-on-one video conversation with another person, and it was only at the end of the call that they were asked if they suspected the other party might not be human. The company stated that participants usually began to suspect something within 20 seconds. Tavus also mentioned that Griffin performed exceptionally well in the NVIDIA's VideofDB benchmark tests, with an average audio and video latency of 0.43 seconds, demonstrating its advantages in real-time conversations.