Tavus launches Griffin, a real-time AI human that tricked 48% of callers

Tavus just previewed Griffin, an AI model that can hold a real-time, face-to-face video call so convincingly that nearly half of early testers thought they were talking to an actual human.

Most conversational avatars today feel like talking to a walkie-talkie with a plastic mask attached. You speak, wait an awkward two seconds, and watch a mouth loosely flap along to robotic speech. Tavus wants to kill that turn-based lag entirely. The startup calls Griffin the first Human Interaction Model (HIM)—an end-to-end, full-duplex video-to-video system that watches, listens, talks, and gestures simultaneously.

In a blind test with 54 participants having one-minute video chats, 48% (26 people) mistook the Griffin-Lite preview for a real person. Tavus says its previous setup, which chained together separate models like Phoenix 4.5, had a pass rate of just 2.4%.

Reading the room in sub-second cycles

Griffin achieves this by ditching the traditional pipeline of speech-to-text, LLM generation, and text-to-speech. Instead, it runs continuous conversational modeling in sub-second loops.

Because it evaluates the exchange constantly, Griffin can nod along while you talk, drop natural verbal cues, or pause when you are just gathering your thoughts rather than handing over the floor. You can interrupt it mid-sentence, and it pivots instantly without losing the thread.

It also brings visual context into the loop. Griffin tracks facial expressions, eye contact, and whatever you share on screen—like spotting a detail in an image and referencing it naturally in dialogue. The model generates full scenes frame-by-frame from a single reference image, including background shadows and ambient posture shifts, while its audio engine can clone a voice from a 10-second clip and stream latency packets down to 10 milliseconds.

On NVIDIA's VideoFDB benchmark, which rates natural conversational video, Griffin-Lite scored 3.83 out of 5 on generation behavior, trailing human reference recordings at 3.92 and beating its nearest rival (Gemini 2.5 plus Anam at 2.80) by over a point. On visual perception, however, the gap was wider: 3.73 versus 4.20 for humans.

Tavus is keeping Griffin-Lite limited to a pool of trusted testers while building out safety and synthetic media disclosure tools, with no firm date for a public rollout.

Why it matters

Tavus took the right approach by treating real-time human nonverbal cues as a foundational challenge rather than a cosmetic skin. If AI agents are going to move into video calls, sub-second latency and visual feedback are mandatory. In education or elder care, an avatar that actually looks at your work, notices confusion, and nods along could turn a clumsy tool into genuinely engaging software.

At the same time, this is an undeniable gift to social engineers if guardrails fail. We have already seen cases like the Hong Kong finance employee who sent $25 million after a deepfake video call with a faux CFO. An avatar that can react on the fly to spontaneous questions makes visual verification practically useless. There is also an asterisk on the testing: 48% fooled in a single sixty-second call with an expectation of meeting a human does not mean it holds up over a twenty-minute meeting. Tavus is smart to keep this locked down, because convincing real-time video generation is a door you cannot close once it opens.