AI Passed Video Turing Test: Tavus Unveiled Griffin Model

Published: October 3, 2026 – 09:26 UTC

Tavus has created an AI that can see, hear, and speak at the same time.
Griffin passed the Turing test and outperformed 15 models in an NVIDIA benchmark.
48% of people did not recognize the AI during a video call.

Tavus has unveiled Griffin — a new-class Human Interaction Model (HIM) that can see, hear, speak, and respond to a conversation partner’s nonverbal cues in real time, all at once. The developers call it the first model to pass the video Turing test.

Introducing Griffin, the first model to pass the video Turing test.

48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.

It’s the first Human Interaction Model (HIM). pic.twitter.com/eb11XbC8Xr
— Tavus (@tavus) October 1, 2026

During testing, 48% of participants in a one-minute video conversation believed they were talking to a real person, while Tavus’ previous system posted a 2.4% result.

For now, the model is available only to a limited group of testers as a research preview. Tavus notes that a number of security issues must be addressed before a public release, as Griffin’s ability to impersonate a human creates additional risks.

Griffin Can See, Hear, and Respond Simultaneously

Griffin is built as a full-duplex video-to-video system. Unlike cascaded systems that first recognize speech, then generate a response, and only then voice it, Griffin makes decisions at regular sub-second intervals.

The model can:

Respond to pauses, gaze, facial expressions, and gestures from the other person
Interrupt or keep the conversation going without losing context
Listen and speak at the same time
See objects or the surrounding environment and factor them into its response
Adjust tone of voice, facial expressions, and gestures to match the situation
Understand the length of a pause and independently choose the right moment to respond

Tavus demonstrated this in practice. Griffin could play Simon Says while simultaneously tracking the words and movements of the people it was talking to. In another demo, the model worked a peacock shown to the camera into the conversation’s storyline. It also helped solve a Rubik’s Cube by following hand movements, and guided soldering work by keeping track of time.

The system consists of two core components: Continuous Conversational Modeling, which analyzes audio and video and determines when and how to respond, and Audio-Visual Generation, which turns those signals into speech and video.

According to Tavus, Griffin-Lite generates 720p video in 320 ms chunks in real time. On H100 GPUs, the average latency was 0.43 seconds — roughly twice as fast as the next-fastest method in the test.

Griffin Outperformed Other Models in NVIDIA Tests

In independent NVIDIA VideoFDB testing designed to evaluate full-duplex audiovisual interaction, Griffin-Lite ranked first among 15 evaluated models.

In the generation track, the model scored 3.83 points versus 3.92 for the human benchmark, and in the perception track — 3.73 points. Griffin also posted the highest alignment score for when the model starts speaking: 62.8% in generation and 73.8% in perception.

*Griffin-Lite’s metrics in the NVIDIA VideoFDB rankings. Source:* *Tavus*.

The most telling was a test with 54 participants. After a one-minute video call, 26 people, or 48%, said their conversation partner was a real person. For the previous system, Phoenix-4.5 combined with Sparrow-2 and Raven-1, that figure was 2.4% — just one out of 41 participants.

At the same time, more than half of Griffin-Lite participants did not think at all during the conversation that they might be talking to an AI. Those who suspected their conversation partner was artificial usually did so within the first 20 seconds.

It was precisely Griffin’s ability to convincingly imitate a human that led to limited access to the model. Tavus said the company is working on safe AI disclosure features and additional risk assessment procedures.

“The same properties that make Human Interaction Models powerful interfaces for natural communications between human and machine allow them to deceive a human into believing it is not AI,” the developers noted.

Tavus is planning a broader release of a more powerful version of Griffin after addressing security issues.

As a reminder:

Anthropic unveiled a cheaper and faster AI model, Claude Opus 5.5
OpenAI unveiled GPT-6 Sol and Luna with lower usage costs
Google released Gemini 4 Argon with a million tokens and unlimited cybersecurity protection

Сообщение AI Passed Video Turing Test: Tavus Unveiled Griffin Model появились сначала на INCRYPTED.

Source: Incrypted