Griffin is Tavus's new full-duplex, video-to-video model that holds face-to-face conversations in real time, continuously seeing, hearing, speaking, moving, and reacting so that what it notices changes what it says and does as a conversation unfolds. It is aimed at developers and product teams building conversational video interfaces, from tutors who can tell when a student is stuck to customer support flows where someone holds a broken part up to the camera. The launch matters because on NVIDIA's VideoFDB benchmark Griffin scored 3.83 out of 5 against a 3.92 human reference and 2.80 for the next-highest published system, and in a separate live study 48% of participants believed they were speaking with a real person rather than an AI.
Tavus is calling this a new category of system, a Human Interaction Model, which bundles perception, conversational timing, speech, and video generation into one model that can react while it is both speaking and listening. Griffin builds on the company's earlier work with Phoenix-4.5 for real-time rendering of subtle active-listening behaviors, Raven-1 for multimodal perception of tone, expression, and visual context, and Sparrow-2 for conversational dynamics like prosody, backchannels, and interruptions. Because the same capabilities that make these models powerful interfaces also make AI harder to distinguish from a person, Tavus is initially releasing Griffin-Lite only to select trusted testers as a research preview while it develops additional disclosure and safety mechanisms.
Tavus is a San Francisco company led by CEO Hassaan Raza, with a research team led by Ioannis Patras and Maja Pantic whose work has been cited over 90,000 times, backed by CRV, Sequoia Capital, Scale Venture Partners and Y Combinator. The company counts Amazon, Mayo Clinic and Salesforce among its customers, alongside 150,000 developers and creators building on its models.
Comments (10)
Sign in to join the discussion.
N
Nadia B.12d ago
48% is wild but also means half the testers caught it. Would love to see the transcripts of what gave it away, that's the actually interesting data.
D
Dhara Mehta12d ago
Curious what the drop-off looks like past the 3-minute mark on live conversations. Uncanny valley usually leaks in once the small talk ends.
C
Chidi Okonkwo12d ago
Rewriting the tagline on the house: 'Griffin. The first AI you forget isn't one.' You can venmo me later.
Z
zoe y.12d ago
HIM is a bold acronym choice. Someone in that naming meeting had conviction and I respect it.
R
Ravi Suresh12d ago
Cool demo, but how are you handling the GPU bill when 10k people hit this live tomorrow? Latency under load is where video avatars usually eat dirt.
M
Mette Lund12d ago
The split-screen reveal in the launch video is the right call, lets the viewer be the judge instead of you telling them. Pacing on the second half dragged a bit though.
H
Herbert Oyelaran12d ago
SOC2? HIPAA? Any posture on biometric data retention? Half my legal team just felt a disturbance in the force reading 'Human Interaction Model.'
M
Marco Pellegrini12d ago
Been tracking video avatar launches all year and this is the first one where the mouth shapes don't look like a ventriloquist dummy having a stroke. Progress.
J
Jun Park12d ago
Naive question: if 48% think it's human, what happens the first time someone asks Griffin to confirm it's not a bot on a sales call? Is there a hardcoded tell?
Y
Yusra Inkwell12d ago
The tweet buries the lede. 'First AI to pass the video Turing test' should've been the whole post, the benchmark line is a footnote.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Tauvus
Introducing Griffin, the first AI to pass the video Turing test
Introducing Griffin, the first AI to pass the video Turing test
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Griffin is Tavus's new full-duplex, video-to-video model that holds face-to-face conversations in real time, continuously seeing, hearing, speaking, moving, and reacting so that what it notices changes what it says and does as a conversation unfolds. It is aimed at developers and product teams building conversational video interfaces, from tutors who can tell when a student is stuck to customer support flows where someone holds a broken part up to the camera. The launch matters because on NVIDIA's VideoFDB benchmark Griffin scored 3.83 out of 5 against a 3.92 human reference and 2.80 for the next-highest published system, and in a separate live study 48% of participants believed they were speaking with a real person rather than an AI. Tavus is calling this a new category of system, a Human Interaction Model, which bundles perception, conversational timing, speech, and video generation into one model that can react while it is both speaking and listening. Griffin builds on the company's earlier work with Phoenix-4.5 for real-time rendering of subtle active-listening behaviors, Raven-1 for multimodal perception of tone, expression, and visual context, and Sparrow-2 for conversational dynamics like prosody, backchannels, and interruptions. Because the same capabilities that make these models powerful interfaces also make AI harder to distinguish from a person, Tavus is initially releasing Griffin-Lite only to select trusted testers as a research preview while it develops additional disclosure and safety mechanisms. Tavus is a San Francisco company led by CEO Hassaan Raza, with a research team led by Ioannis Patras and Maja Pantic whose work has been cited over 90,000 times, backed by CRV, Sequoia Capital, Scale Venture Partners and Y Combinator. The company counts Amazon, Mayo Clinic and Salesforce among its customers, alongside 150,000 developers and creators building on its models.
Comments (10)
48% is wild but also means half the testers caught it. Would love to see the transcripts of what gave it away, that's the actually interesting data.
Curious what the drop-off looks like past the 3-minute mark on live conversations. Uncanny valley usually leaks in once the small talk ends.
Rewriting the tagline on the house: 'Griffin. The first AI you forget isn't one.' You can venmo me later.
HIM is a bold acronym choice. Someone in that naming meeting had conviction and I respect it.
Cool demo, but how are you handling the GPU bill when 10k people hit this live tomorrow? Latency under load is where video avatars usually eat dirt.
The split-screen reveal in the launch video is the right call, lets the viewer be the judge instead of you telling them. Pacing on the second half dragged a bit though.
SOC2? HIPAA? Any posture on biometric data retention? Half my legal team just felt a disturbance in the force reading 'Human Interaction Model.'
Been tracking video avatar launches all year and this is the first one where the mouth shapes don't look like a ventriloquist dummy having a stroke. Progress.
Naive question: if 48% think it's human, what happens the first time someone asks Griffin to confirm it's not a bot on a sales call? Is there a hardcoded tell?
The tweet buries the lede. 'First AI to pass the video Turing test' should've been the whole post, the benchmark line is a footnote.