Griffin scores 3.83 against NVIDIA’s 3.92 human benchmark, the highest score ever recorded for an AI, while in blind testing, 48% of participants thought they were talking with a real person

Tavus, the human computing company, today announced Griffin, the first Human Interaction Model, creating a new class of AI built to understand and participate in human interaction. Griffin is a full-duplex, video-to-video model that holds face-to-face conversations in real time. It sees, hears, interprets, speaks, moves, reacts, and responds continuously, allowing what it notices to change what it says and does as a conversation unfolds.

This press release features multimedia. View the full release here: https://www.businesswire.com/news/home/20261001092598/en/

AI has advanced rapidly, but people still have to communicate on the computer’s terms: through text, commands, or voice. Human Interaction Models flip that relationship, enabling machines to interact the way people already do.

Griffin is uniquely able to process the full scale of face-to-face human interactions. It can adapt words, timing, tone, expressions, and movements to what is happening in the moment, just as a live person would. Think reacting to interruptions, recognizing when to laugh or lean in, or waving at someone who just entered the room. Griffin responds dynamically as conversations unfold, even if a conversation takes an unexpected turn.

This groundbreaking model is stunningly life-like and is already approaching human performance in independent evaluations of face-to-face conversational AI. On NVIDIA’s VideoFDB benchmark, Griffin scored 3.83 out of 5 on generation, compared with a 3.92 human reference and 2.80 for the next-highest published system. In a separate live study, 48% of participants believed they were speaking with a real person rather than an AI.

“For decades, we’ve imagined computers as partners, not just tools,” said Hassaan Raza, CEO of Tavus. ”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. That’s what we mean by human computing.”

Built on Industry-leading Research

Griffin’s breakthrough was made possible by years of Tavus research into the different components of human interaction. Tavus’ research team, whose work has been cited more than 90,000 times, has developed industry-leading models across perception, conversational timing, and expressive video. More than 150,000 developers and businesses already use one or more of these technologies.

Phoenix-4.5 is Tavus’ industry leading real-time rendering model, designed to generate the subtle movements and active-listening behaviors that make face-to-face interaction feel natural. Raven-1 focuses on multimodal perception, allowing AI to understand not only what someone says, but signals including tone, expression, and visual context. Sparrow-2 focuses on the dynamics of conversation, modeling signals including prosody, backchannels, interruptions, and the broader audio environment.

Griffin builds on the research behind these systems but represents a fundamentally different approach: a single Human Interaction Model designed around the interaction itself. Perception, conversational decision-making, speech, and video generation operate continuously, allowing the model to react while it is speaking as well as while it is listening.

The result is a model that does more than process a conversation. It participates in one.

Griffin in Action

To get a sense of what Griffin is, Tavus is releasing a short film, which features Vanessa, one of Tavus’ PALs– the human representation that interacts with users. Vanessa runs on top of Griffin. She is the most realistic PAL ever built.

That could mean a tutor who notices when a student is struggling before they know how to ask for help, an employee rehearsing a difficult conversation with someone who reacts dynamically, or a customer holding a broken part up to the camera and working through a solution without needing to know what the part is called. In each case, the person can simply communicate as they naturally would, while the model works to understand what they mean from their words, behavior, and surroundings.

Because the same capabilities that make Human Interaction Models powerful interfaces can also make AI more difficult to distinguish from a person, Tavus is initially making Griffin-Lite available only to select trusted testers as a research preview. The company is developing additional disclosure and safety mechanisms before making Griffin broadly available.

To meet Vanessa and see Griffin in action, watch the short film here. For updates and additional information, please visit www.tavus.io.

About Tavus

Tavus is a San Francisco-based AI research company pioneering human computing: the pursuit of teaching machines the art of being human so that interacting with a machine can feel as natural as interacting with a friend or coworker. Tavus trains the foundational models that power every PAL, or human representation. The work of its research team, led by Ioannis Patras and Maja Pantic, has been cited over 90,000 times. The company has raised $70M in funding and is backed by CRV, Sequoia Capital, Scale Venture Partners and Y Combinator. Customers include leading enterprises such as Amazon, Mayo Clinic and Salesforce, as well as 150,000 developers and creators. Learn more at www.tavus.io.

View source version on businesswire.com: https://www.businesswire.com/news/home/20261001092598/en/

”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."