news

Tavus says Griffin-Lite passed a video Turing test, with 48% of 54 testers fooled

The October 1 research preview puts perception, turn-taking and video generation in one real-time model, and the headline number comes from a study Tavus ran on itself.

By Priya Shenoy · 1 October 2026 · Updated 4 October 2026

Tavus announced Griffin on October 1, 2026 and attached a single headline number to it: in a company-run study, 26 of 54 people who held a one-minute video call with a Griffin-Lite persona came away believing they had talked to a real person. The comparison point Tavus gives for its own previous system is 1 of 41, or 2.4%. Griffin-Lite is a research preview limited to selected testers, and the company says the more capable Griffin will follow once it works out how to release it safely.

One model handles perception, timing and video generation together

Tavus calls Griffin a Human Interaction Model, which it describes as a new class of model built to understand and generate face-to-face real-time interaction. The practical difference from the avatar stacks that preceded it is architectural. Earlier systems chained speech recognition, a language model, speech synthesis and a face animator, and each link in that chain waited for the caller to stop talking. Griffin is full-duplex video-to-video, so perception, the decision about when and how to respond, and the generation of speech and video all run at once. Tavus says the system renders the full scene from a single reference image, including body movement and the surroundings, rather than animating a face over a fixed plate.

The 48% figure is a vendor study, not an independent evaluation

The Turing test framing is doing a lot of work here, and it is worth being precise about what was measured. The sample is 54 people. The interaction is one minute long. The design, the personas, the call conditions and the scoring all sit with the company making the claim. That does not make the result fake, and a jump from 2.4% to 48% on comparable internal setups is a large enough delta that methodology noise is unlikely to explain all of it. It does mean the number is not yet reproducible by anyone outside Tavus, which is the same problem that shows up every time a vendor grades its own output, as with Tencent's self-scored image win rate.

NVIDIA built and scored the benchmark Tavus leans on

The more interesting evaluation detail is buried in the acknowledgements. Tavus credits NVIDIA with building and scoring the Video Full-Duplex Benchmark, and Queen Mary University of London with research collaboration, alongside Baseten, Daily and Cerebrium for preview infrastructure. A benchmark for full-duplex video, scored by a third party rather than the model vendor, is the kind of measurement this category has largely lacked. Avatar and talking-head launches usually ship with demo reels and adjectives, as when Google shipped Live Avatar without comparative quality numbers.

Griffin is not available to buy, and Tavus names the reason

Griffin is not on the Tavus platform. The company says it will arrive once safe release is worked out, and points existing customers to Phoenix, Raven and Sparrow in the meantime. Tavus states the risk plainly in its own post, writing that the same properties that make these models good interfaces also "allow them to deceive a human into believing it is not AI." That restraint is easier to credit when you note what Tavus already sells: a platform where AI personas join calendar-invited meetings on Zoom, Meet and Teams. The capability gap between the research preview and the shipping product is the only thing holding the two apart.

Sources: Tavus, Cybernews.