Tavus's Griffin video model passed for human with 26 of 54 people in a company-run test

Because the partner was billed as a fellow participant, the setup leaned answers toward "human" before anyone spoke ^[3]. Even so, 28 of the 54 did not take Griffin for a person ^[2]. The Phoenix-4.5 arm controls for that lean, since the same cover story produced almost no human verdicts for the older system ^[4]. The published account does not say whether participants were assigned to the two arms at random or tested in the same period.

Fifty-four people is a small sample. A normal-approximation 95% interval on 26 of 54 runs from about 35% to 62% ^[3]. RuntimeWire, reporting the announcement, describes it as a small company-run result and not a broad measure of how often people would mistake Griffin for a human in ordinary use ^[13]. For the number to transfer, a deployment would have to look like the test: one minute, small talk about the year ahead, and a caller with no reason to suspect software. Tavus lists tutoring, practice for difficult conversations and camera-based tech support as uses ^[16]. Each of those runs past a minute and keeps the model on a task.

The design target is the right one. Griffin takes in audio and video continuously and decides whether to speak, wait or yield, producing voice and video together ^[9]. Building the model around turn-taking is sound engineering, because turn-taking is what a one-minute call exposes first. Tavus says it renders full scenes and responds to objects held up to the camera, and one demo has it playing Simon Says ^[10]. That is a fair turn-taking test, if not one a buyer will write into a contract.

The latency figure covers one component. Tavus says the video-generation stage answers incoming audio with a frame in an average of 0.43 seconds on NVIDIA H100s, about half the latency of the next-fastest published system it compared against ^[11]. A caller notices the slowest replies, and a mean on top-end GPUs describes the typical one.

The NVIDIA scores come from a language-model judge, and RuntimeWire notes they measure something different from the live study ^[8]. Tavus describes the benchmark run as independent, according to The Decoder ^[14]. Griffin-Lite scored 3.83 of 5 on generation against a human reference of 3.92, and 3.73 on perception against 4.20 ^[6][7]. That leaves it 0.09 points short of human on generation and 0.47 short on perception ^[4][5]. Its lead over the next system is 1.03 points on generation and 0.29 on perception ^[6][7].

Tavus is withholding Griffin because the realism that makes it useful can also make it deceptive ^[12]. A more capable version will follow once "safety concerns are addressed," the company says ^[15]. The disclosure gate is Tavus's own decision. I think it is the right one for a company whose own test had about half its participants fooled inside a minute ^[1]. Tavus raised a $40 million Series B in November 2025, led by CRV ^[17], and about $64 million in total, according to The Decoder ^[18].