Tavus Launches Griffin, a Video Model Built to Read and Return Facial Cues in Real Time
Tavus says Griffin is the first conversational AI model designed for face-to-face interaction, reading a user's expressions and responding with its own on-screen face.
Tavus has unveiled Griffin, a model the company describes as the first built specifically for face-to-face human interaction, moving conversational AI past voice and text into a format where both sides see and react to a face. The announcement, made through Business Wire, positions Griffin as an interaction layer rather than a content-generation tool: it is meant to run inside live conversations, not produce clips after the fact.
Most video avatars on the market today generate a talking head from a script, then stop. Griffin's pitch is the reverse loop — the model takes in a camera feed, picks up on the user's facial signals, and adjusts its own expression and delivery turn by turn. That shift matters because turn-taking in a real conversation depends on cues that arrive in fractions of a second, and a system that only speaks cannot participate in them.
Tavus has not yet published the underlying architecture, benchmark results, or latency figures in the announcement, and those numbers will decide whether Griffin feels like a conversation or a demo. Real-time expression tracking typically requires low-latency inference on video, a workload far heavier than text generation. Whether Tavus runs this on-device, in the cloud, or in a hybrid setup is one of the first questions technical buyers will ask.
The company's existing business is built around personalized video avatars, used by sales and marketing teams to produce one-to-many clips at scale. Griffin pushes that stack toward one-to-one sessions — onboarding, support, tutoring, or any scenario where a scripted video will not do. It also puts Tavus closer to the crowded digital-human space occupied by firms such as Synthesia, HeyGen, and several startups chasing real-time agents.
The immediate commercial question is who needs a face on their AI at all. Voice agents are cheaper to run and already handle routine support calls; adding video raises bandwidth, compute, and privacy costs, since a camera feed of a user's face is sensitive data. Tavus will need to show that the facial back-and-forth measurably improves outcomes — completion rates, trust, or sales — rather than just looking better in a launch video.
What to watch next is pricing and access. Tavus has not said when Griffin will be generally available, what it costs, or which developers get it first, and those details will shape whether this becomes an API other products build on or a showcase feature inside Tavus's own tools. If the latency holds up under real network conditions, the more interesting test is whether users forget they are on a call with a model — and how they feel once they remember.