Forget chat interfaces and voice bots. Tavus just announced Griffin, a video-to-video AI model that holds real-time face-to-face conversations, and nearly half of people tested couldn’t tell it wasn’t human. That’s not a marketing claim. It’s from a blind study.
Griffin is a full-duplex model, meaning it sees, hears, speaks, and reacts all at once, continuously. It doesn’t wait for you to finish talking. It notices when you laugh, when you hesitate, when someone walks into the room behind you. Tavus calls this category a ‘Human Interaction Model,’ positioning it as a new class of AI distinct from existing voice or avatar products.
How it scores against human performance
On NVIDIA’s VideoFDB benchmark, Griffin scored 3.83 out of 5 on generation quality. The human reference score is 3.92. The next-best published system sits at 2.80. That’s a significant gap, and Griffin is sitting just below human-level performance on an independent benchmark, not an internal one.
Griffin is built on three existing Tavus models: Phoenix-4.5 for real-time video rendering, Raven-1 for multimodal perception, and Sparrow-2 for conversational dynamics like interruptions and backchannels. But Griffin isn’t just a combination of those. It’s designed around the interaction itself, with perception, speech, and video generation running simultaneously rather than sequentially.
Who should pay attention
Competitors in the video avatar space, including HeyGen and Synthesia, have focused on scripted or asynchronous video. Griffin is targeting live, reactive conversation, which is a harder problem and a different market. Real use cases here include:
- AI tutors that detect confusion without being asked
- Sales or HR roleplay tools with dynamic responses
- Customer support where users describe problems visually, in real time
Still, Tavus is rolling out carefully. Griffin-Lite is available to select testers only, with broader access gated behind additional safety work. Given how close to human-level the model is, that caution is probably the right call.
