Tavus Unveils Griffin, a Real-Time Face-to-Face Conversation Model: Not Yet Available

AI video company Tavus unveiled Griffin, a new class of model it calls a "Human Interaction Model." In a video call, Griffin hears and sees you and responds in real time with a face that talks and moves. The company presents it as the next step, but it isn't releasing the model yet.

In brief

• What was announced: Griffin, on October 1. A video conversation model that can listen and talk at the same time ("full-duplex"), perceives both audio and video, and generates the whole scene, including the face, body, background, and shadows. The company says it can clone a voice from about 10 seconds of audio.
• The company's technical claims: about 0.43 seconds average latency and 720p video. It says it ranked first among 15 models on a video conversation benchmark scored by NVIDIA. These are Tavus's own reported results.
• The "real person?" study: in Tavus's own experiment, 26 of 54 participants who spoke with Griffin's preview version believed they were talking to a real person. With the company's previous system, it was 1 of 41. Tavus presents this as "the first model to pass the real-time, video Turing test." That's the company's own claim, and the sample is small.
• Status: a research preview open only to select trusted testers. No pricing. The company says it plans to release it once safety concerns are addressed.

Our take

The company itself wrote the most honest line in this story: the same qualities that make these models a natural interface between people and machines also make it possible for them to deceive. Tavus says that's why it isn't releasing Griffin yet, and that it's working on disclosure features that make clear you're talking to an AI. For our audience, there's no tool to use today. But there are two practical takeaways for the near future. First, in a video call, someone's face and voice are steadily becoming less of a guarantee that it's really them. It makes sense to build the habit of confirming through a second channel, especially in video calls that ask for payment, passwords, or sensitive information. Second, once this technology matures, it will show up in customer service and sales calls too. If you use it, telling your customers clearly that they're speaking with an AI is how you manage both trust and legal risk. The "passed the test" claim also rests on the company's own small study for now; let's wait for independent results.

Kaynak: Tavus