Tavus Griffin: what it is and who can use it

Tavus Griffin is a real-time video model that holds a face-to-face conversation: it watches you on camera, listens, and talks back with a generated face and voice, all at once. Right now almost nobody can use it. The only version out is Griffin-Lite, a research preview for a short list of trusted testers, with no public API and no price.

The post that got 20 million views

On October 1, @tavus announced Griffin as "the first model to pass the video Turing test." The headline number: 48% of people who talked to it live thought it was a real person, against under 3% for earlier systems. Tavus calls it a "Human Interaction Model," or HIM. By the next day the post had passed 20 million views.

A follow-up clip from @tavus shows the part that makes it different. A man named Ari works a Rubik's cube while Griffin coaches him from what it sees in his hands, and when he goes quiet to think, it waits. Tavus says the model's perception is always on, so it can act without being asked.

@stephzhan quoted the launch and pointed to the same behaviors: noticing an object you hold up, waiting while you think, laughing and cutting in while you are still talking. Her read is that future AI interfaces will feel less like prompting a machine and more like talking to a person.

What Griffin actually does

Most voice and avatar bots take turns. You talk, they wait for silence, they answer. Griffin is full-duplex: it takes in video and audio and produces speech and video at the same time, several times a second. That lets it nod or say "mm-hm" mid-sentence, stop when you interrupt, and react to things in the frame. Decrypt reported an average audio-to-video delay of 0.43 seconds on Nvidia H100 hardware.

Tavus lists three uses: tutoring students who are stuck, letting employees rehearse hard conversations, and customer support where the AI can identify a part the customer holds up without the customer knowing its name.

How solid is the 48%?

It's shakier than the headline suggests. X readers attached a Community Note to the launch post saying the figure comes from Tavus's own study of 54 one-minute calls, with no independent check and no standard protocol. According to Decrypt, participants were told they would be matched with another person and only asked afterward whether they suspected AI. 26 of 54 said they thought it was human. The older system fooled 1 of 41.

So it's a small test, run by the vendor, on calls that lasted a minute, and telling people they're about to meet a human primes them to believe it. I'd still take it as a sign the jump from the old system is real, just not as proof that Griffin passes for a person in general.

The benchmark claim holds up better. On VideoFDB, a full-duplex video benchmark run by Nvidia, Griffin-Lite scored 3.83 out of 5 on generation. The human reference scored 3.92 and the next-best model 2.80. On perception it scored 3.73 against 4.20 for humans, so the gap there is wider.

Is Tavus Griffin free, and can you get it?

No to both, for now. Griffin-Lite is a research preview, and the only way in is a request form on the Tavus site. Tavus hasn't published pricing or a release date. The company's stated reason for holding back is that a model people mistake for a person needs disclosure and safety work before a wide release. Vendors rarely say that out loud, and I think they're right to wait.

What this means for a business

Tavus hasn't given a date, but video bots that can hold a natural conversation look close to something a tutoring company or a support desk could buy. Once they can, the question for the owner is less about quality and more about whether customers know they're talking to software.

We give clients the same advice for voice and chat today: say it's AI up front, every time. Customers who find out afterward feel tricked, and that costs more trust than the bot saved in staff time. The realism is useful for making a conversation smoother. It shouldn't be used to hide what's on the other end.

For most small businesses, the more practical work sits after the conversation: what happens once the call ends, who gets the summary, and how the follow-up goes out. If you want help mapping where a conversational AI fits in your intake or support, and where it shouldn't, New Face Design's free process audit is a good place to start.