Breaking

Tavus Says Griffin Reached 48% in a One-Minute Video Human-Likeness Test

October 4, 2026 at 13:45 EDT

Foundation Models
AI Agents
Media Generation

Tavus introduced Griffin, a real-time video-to-video AI model that the company says made 48% of participants believe they were speaking with a human after a one-minute live video call. The San Francisco company is calling Griffin its first Human Interaction Model, a system designed to see, hear, speak and generate expressive video at the same time rather than passing a conversation through separate speech, language and avatar components.

October 2026 · Tavus

Griffin tests the edge of live video AI

In one-minute video calls, Tavus says nearly half of participants believed Griffin-Lite was a real person, compared with 2.4% for its previous system.

48%

Believed Griffin-Lite was human

2.4%

Previous Tavus system result

Participants in Griffin-Lite test

0.43s

Average audio-to-video latency on H100s

Human-likeness result, drawn to scale

2.4%Previous system

48%Griffin-Lite

Each orange segment represents the same share as the blue segment. The test result was about 20 times higher than Tavus's previous system.

What changes in the model loop

Sees and hears

Incoming video and audio stay active during conversation.

→

Decides timing

The system can wait, interrupt, backchannel or respond.

→

Generates video

Speech, face, body and scene are generated in real time.

Why it matters

Developers see possible uses in tutoring, coaching, customer support and repair guidance where visual context and timing matter.

What remains unresolved

The preview is limited to trusted testers, and concerns center on disclosure, deception risk and the short test window.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.