Tavus Releases Griffin-Lite: First Model to Pass a 'Video Turing Test'?

Tavus launched Griffin, which the company calls the first "Human Interaction Model" (HIM), a class of models designed to understand and conduct face-to-face conversation in real time. This positioning was put forward by Tavus itself, and the report relays it with "the company says," without providing third-party recognition.

Released alongside the positioning was a set of highly striking performance data. A Tavus study claims that 48% of participants believed Griffin was a real person after a one-minute video call; the same study claims that previous systems topped out at only 2%.

These two sets of numbers—48% and 2%—both come from Tavus's in-house research, and the report provides no independent verification. Therefore, they can only be treated as vendor claims, not as facts confirmed by a third party.

In addition to its in-house research, Tavus also relays what it calls an independent test conducted by Nvidia, which measures how "human-like" AI is in direct audio-video conversation. In this test, Griffin scored 3.83, a real person scored 3.92, and the previous best AI model scored 2.80.

These numbers can likewise only be treated as vendor claims. The report does not disclose the specific design, sample, or scoring criteria of the test, and its degree of "independence" cannot be verified. The conclusion of the so-called "independent test" entered the report through Tavus's relay, rather than from a first-party Nvidia release.

In other words, from the 48% in Tavus's in-house research to the 3.83 in the relayed Nvidia test, the entire evidence chain originates on the vendor's side. The report provides no independent third-party verification of any of these numbers.

According to the report, Griffin processes speech, facial expressions, tone, gestures, and pauses while receiving and generating video. This is the most specific description of Griffin's capabilities currently available, and it is a confirmed product feature statement.

However, the report does not state Griffin's pricing, compute requirements, latency metrics, or commercial timeline. For a model that emphasizes "real-time face-to-face conversation," these omissions mean outsiders cannot assess its actual deployment conditions.

More details about Griffin, the report says, are contained in Tavus's research report. This means that more complete technical information was not made public through the report itself.

The currently accessible version is called Griffin-Lite, available as a "research preview" to "select testers." This is not a publicly available product, but a preview version with limited capabilities.

The version with more complete capabilities will be released after "safety concerns are resolved." This wording confirms that the full version has not yet been released, but the report does not explain what the "safety concerns" specifically refer to, who determines them, or when they count as resolved.

From this it can be inferred that outsiders can currently only access the capability-limited Griffin-Lite research preview, and the capability boundaries and release timing of the full version remain undetermined. This inference is based on the conditional wording "safety concerns are resolved" in the report, rather than any published timeline.

Tavus is a San Francisco AI startup founded in 2020, which has raised about $64 million. The company initially produced personalized AI videos for sales and marketing, then expanded to real-time video conversations with digital humans.

Tavus's listed potential use cases include tutoring, practicing difficult conversations, and camera-based technical support. These are potential directions listed by the company, and the report does not verify Griffin's actual performance in these scenarios.

From marketing videos to real-time digital human conversation, Tavus's business evolution direction is consistent with Griffin's positioning. But the funding scale and founding date themselves do not constitute independent endorsement of Griffin's performance data.

Confirmed facts include: Tavus launched Griffin; Griffin processes speech, expressions, tone, gestures, and pauses and simultaneously receives and generates video; Griffin-Lite is available as a research preview to select testers; the more complete version will be released after safety concerns are resolved; Tavus was founded in 2020, has raised about $64 million, initially made marketing videos and later shifted to real-time digital human conversation; potential use cases include tutoring, difficult conversation practice, and camera-based technical support.

Vendor claims include: the 48% misidentification rate, the previous systems' 2% ceiling, and the three scores of 3.83, 3.92, and 2.80 in the relayed Nvidia test. These numbers all come from Tavus's in-house research or its relay, and the report provides no independent verification.

Inferences include: "the first Human Interaction Model" is Tavus's self-proclamation, with no third-party recognition; the key numbers can only be treated as vendor claims due to the lack of independent verification; the full version has not yet been released because safety concerns remain unresolved, leaving outsiders able to access only the capability-limited preview.

The dimensions not disclosed by the report also constitute the boundaries of evidence: the test sample size, participant composition, test environment, and scoring method are all undisclosed; whether the 48% misidentification rate can be reproduced in a larger-scale, independently designed experiment remains unclear; how much more capable the full version is compared with Griffin-Lite is not quantified in any way in the report; and the risks and safeguards against misuse of the technology for impersonation, fraud, and similar scenarios are not addressed in the report.