Tavus previews Griffin, a full-duplex video model for face-to-face AI conversation

Tavus announced Griffin on 1 October 2026, describing it as "the first Human Interaction Model" — a full-duplex video-to-video system built for real-time face-to-face conversation. The company says the model sees, hears, interprets, speaks, moves and reacts continuously, adapting words, timing, tone, expressions and movement during an interaction, handling interruptions, and responding to social cues such as laughing or leaning in and to changes in the environment.

Tavus reports that Griffin scored 3.83 out of 5 on NVIDIA's VideoFDB benchmark, against a human reference of 3.92 and a next-highest system at 2.80, and that in blind testing 48% of participants believed they were talking to a real person rather than an AI. A smaller model, Griffin-Lite, is available as a research preview to "select trusted testers only"; the company says it is developing disclosure and safety mechanisms before a broader release. No pricing was given.

The announcement states that Tavus is based in San Francisco, has raised $70M from CRV, Sequoia Capital, Scale Venture Partners and Y Combinator, and serves more than 150,000 developers and businesses, with enterprise customers including Amazon, Mayo Clinic and Salesforce. Its research team is led by Ioannis Patras and Maja Pantic. Chief executive Hassaan Raza said: "AI has become incredibly intelligent, but we still have to meet the machine on its terms."

Why it matters for digital humans: Full duplex — reacting while listening, not just after a turn ends — is the gap between a talking-head player and something that feels like a conversation, and it is where most avatar stacks still break down. Treat the numbers with care: the benchmark score and the 48% blind-test figure are both company-reported in a press release, the test protocol and sample size are not described, and nothing here has been independently reproduced. The access model is the more concrete signal. A research preview limited to trusted testers, with disclosure mechanisms still being built, means this is not something you can specify into a deployment this quarter, and the safety framing suggests Tavus expects the realism itself to be the regulatory problem.