Tavus says Griffin passes a 48% human test

Tavus says Griffin-Lite, its full-duplex video-to-video Human Interaction Model, convinced 48% of participants (26 of 54) that they were speaking with a person after a one-minute call, according to the company’s Griffin announcement. The release is only a research preview for selected early testers—not a public customer API—and Tavus reports 0.43-second audio-to-video latency on H100s, so the headline is a perception result, not proof of general conversational reliability. Product teams evaluating real-time avatars should file it beside Microsoft’s streaming speech cost boundary and test disclosure, latency, and user trust separately before treating human likeness as product quality.