An AI Passed a Video Turing Test. Nobody at the Company Is Celebrating.

Published: October 3, 2026 | Category: AI | By Mahesh

Today's new model is Griffin, which San Francisco-based Tavus introduced on October 1, 2026 as what it calls the first Human Interaction Model, a full-duplex, video-to-video system that holds a real-time, face-to-face conversation rather than taking turns like a voice assistant [1]. Tavus CEO Hassaan Raza framed the shift directly: "AI has become incredibly intelligent, but we still have to meet the machine on its terms. Griffin is a step toward changing that, toward machines that understand how we naturally communicate and meet us where we are" [2].

The headline result is the video Turing test. In a blind study, 54 participants had a one-minute video call expecting to speak with another study participant; 26 of them, 48 percent, came away believing they had talked to a real person [3]. Tavus ran the identical test on its previous conversational stack, Phoenix-4.5, Sparrow-2 and Raven-1, and only 1 of 41 participants, 2.4 percent, was fooled [4]. That is not an incremental gain. It is close to a twentyfold jump in how convincingly a system can pass as human over a short, unscripted conversation, and it happened in a single model generation.

"Griffin is not on the Tavus platform yet, and will come once Tavus has worked out how to release it safely."

Tavus, closing line of its own Griffin announcement [5]

What makes this launch worth covering is not just the number, it is what Tavus chose to do with it. The company is not shipping Griffin to customers. Griffin-Lite, the version tested, is available only to a small group of trusted testers as a research preview, and Tavus has given no firm public date for a wider release [6]. The company's own stated reason is unusually direct for a product announcement: "because the same capabilities that make Human Interaction Models powerful interfaces can also make AI more difficult to distinguish from a person," the company is deliberately withholding broader access while it works through safety and disclosure measures first [7].

That restraint is a genuinely different posture than most AI launches this year, where capability and availability usually ship together. A company that just demonstrated a result this striking would normally be racing to monetize it. Tavus instead published the number, then paired it with an explicit statement that the result itself is the reason for caution, not the reason to launch faster. Participants in the same study rated Griffin 5.4 out of 7 for seeming natural and 5.6 for seeming trustworthy, scores high enough that an unannounced deployment, a customer service line, a dating app profile, a political robocall, could plausibly go unnoticed by the person on the other end [8].

On the technical side, Griffin also topped Nvidia's independently run VideoFDB benchmark for full-duplex audiovisual conversation, placing first of 15 tested systems on both the generation and perception tracks, scoring 3.83 against a human reference of 3.92 on generation [9]. Tavus says Griffin continuously decides whether to speak, wait or interject within sub-second intervals while processing both audio and video simultaneously, rather than waiting for a pause the way a voice assistant does, and that voice cloning now requires only about ten seconds of reference audio [10]. Tavus already serves roughly 150,000 developers and businesses through its existing Phoenix, Raven and Sparrow models, giving it a real commercial base to eventually deploy Griffin into once the company decides it is ready [11].

What to Watch Next

The real story over the next few months is what "safely" ends up meaning in practice: whether Tavus ships Griffin with a mandatory, unremovable on-screen disclosure that a user is talking to an AI, or relies on something softer like terms of service and customer-facing policy. A twentyfold jump in how convincingly a system passes as human is exactly the kind of capability that regulators, not just the company that built it, are likely to have opinions about once a wider release date actually gets set.

Sources

1. Business Wire via Yahoo Finance, "Tavus Introduces Griffin, the First Face-to-Face Human Interaction Model," October 1, 2026. finance.yahoo.com
2. Business Wire via Yahoo Finance, "Tavus Introduces Griffin," Hassaan Raza quote (see source 1).
3. Dealroom, "Tavus Launches Griffin, a Real-Time Video Model That Passed a Video Turing Test," October 2026. app.dealroom.co
4. Dealroom, "Tavus Launches Griffin" (see source 3).
5. Orca Router, "Tavus Griffin: The 48% That Passed the Video Turing Test," October 2, 2026. orcarouter.ai
6. Complete AI Training, "Tavus Introduces Griffin, a Model It Says Passes the Video Turing Test," October 2026. completeaitraining.com
7. BusinessWire, "Tavus Introduces Griffin, the First Face-to-Face Human Interaction Model," October 1, 2026. businesswire.com
8. BusinessWire, "Tavus Introduces Griffin," safety rationale (see source 6).
9. Dealroom, "Tavus Launches Griffin," naturalness and trust scores (see source 3).
10. Crypto Briefing, "Tavus Unveils Griffin, Claiming the First Video Turing Test Pass," October 2026. cryptobriefing.com
11. The Rundown, "Tavus Previews Griffin, a Lifelike AI That Reacts on Video Calls," October 1, 2026. therundown.ai
12. Dealroom, "Tavus Launches Griffin," existing customer base (see source 3).

Read More on Depth Grid

• OpenAI Shelved Its Next Model Over Lying. Then Shipped Agents on the Old One Anyway.
• Meta's New AI Agent Wants to Book Your Travel and Bills
• What an LLM Actually Is, Explained Through the Two Models OpenAI Just Shipped
• A Slack Rival Where Your AI Agent Has Its Own Inbox, Not Just a Login
• OpenAI Just Refused to Ship Its Own Flagship Model

Article by Mahesh | Depth Grid