MODEL PROFILE · Tavus

Tavus Griffin-Lite

A research-preview video interaction model. Read the access restriction and the limits of its reported one-minute study before treating the announcement as a product launch.

Source reviewed: 2026-10-02 · Video model · restricted research preview

Provider documentation summary. No hands-on results or independent ranking are claimed.

Documented facts

The following fields come from the official provider source, reviewed on the date above.

Identity and announcement: Griffin-Lite is the preview named in Tavus’s October 1, 2026 Griffin announcement.
Inputs and outputs: Tavus describes concurrent audio/video perception and speech/video generation: it can listen while speaking.
Access: Selected trusted testers only. The announcement says Griffin-Lite is not available to customers or on the Tavus platform yet.
Claimed strengths: Continuous turn-taking, visual reactions and expressive scene generation are vendor-described capabilities, not DualView measurements.
API, pricing and weights: No public Griffin endpoint identifier, billing rate or downloadable-weight license verified here. Existing Tavus product plans are not Griffin prices.
Other unknowns: Public release date, session limit, context window and deployment requirements remain unverified.

What the reported face-to-face result measures

Tavus reports that 26 of 54 participants believed their partner was human after a one-minute Griffin-Lite call. Participants expected another participant; disclosure came afterward. The prior Phoenix-4.5/Sparrow-2/Raven-1 system returned 1 of 41 under the reported protocol. These are vendor-reported study results, not our test or proof of general human equivalence. Tavus study methodology and sample counts.

The practical decision

Decide whether the task requires a live exchange or a finished clip. A conversation must respond to something that happens after it starts; a prerecorded presentation can be judged as a complete artifact. Those are different acceptance criteria, so an attractive short demonstration alone cannot establish readiness for an interactive application. If you are evaluating this category, write down the moments that should change the response: a correction, a pause, an object entering the camera view, or a request to stop. Then specify what a successful response would look and sound like before watching the output.

What to record in your own comparison

Use another real-time audiovisual interaction system as the comparator. Give both the same sequence of events and disclose differences in the input streams. Do not convert a comparison with a text-only chatbot into a video-generation quality ranking.
Separate realism from usefulness. Record whether the answer is correct, whether it refers to the right visible object and whether the conversation recovers after a misunderstanding. A convincing face does not settle those questions.
Keep the full session when consent permits, including awkward silences and failed interruptions. A highlight reel cannot show how frequently a problem occurs. Note the client, connection conditions and recording method before comparing timing.
For your own user study, disclose AI use and define the task clearly. Measure successful task completion and appropriate response timing rather than making mistaken identity the product objective.

Limitations and unknowns

This is documentation research. We made no calls, produced no examples and performed no independent performance assessment. The cited study is brief and concerns a particular participant setup; generalization to longer conversations, different participants or a production environment requires separate evidence. The profile records the named preview, and does not assign its reported results to a future larger Griffin version. A public announcement does not provide an integration contract.

Compare your saved results

Use the DualView editor to inspect files you already have. The accepted-output calculator uses your own rates and counts. Neither link runs a model or purchases generation.

Related reporting

Inspect saved video results in the editor

Related model profiles

All model profiles and dated changes · Subscribe via RSS