Tavus Griffin: What Its Benchmarks Mean for Video Agent Testing

What Is Tavus Griffin?

Tavus Griffin is a full-duplex, video-to-video model that Tavus announced on 1 October 2026, in a post by co-founder and CEO Hassaan Raza and head of research Ioannis Patras. Tavus defines a Human Interaction Model as one designed to understand and generate face-to-face, real-time human interaction: it listens while it talks and attends to expressions and pauses as well as words.

Most real-time video agents work as a cascade: speech recognition transcribes you, a language model replies, and separate systems turn the reply into a voice and a face. Every handoff adds delay and drops signals, such as your tone or what is on camera, that the next system never sees. Griffin folds perception, turn decisions, and speech and video generation into one system that runs them all at once.

Tavus's announcement shows these capabilities in recorded conversations with people who had never used Griffin:

• Behavior and emotion - it laughs, changes its tone, and shifts its expression and gestures in response to what the other person says and does.
• Full-scene generation - it generates every pixel of every frame from one reference image, including hands, the chair, shadows, and the background.
• Full-duplex conversation - it can interrupt, back-channel with an "mm-hm", or stop the moment you cut in, without losing its place.
• Perception - it uses what it sees, for example coaching someone through a Rubik's Cube while watching the cube turn.
• Temporal understanding - it tracks how long a silence has lasted and speaks up when the next step of a task is due.

Griffin's Two-Engine Architecture

Griffin splits the work between two engines that run concurrently for the whole conversation:

• Continuous Conversational Modeling - perceives the incoming audio and video, decides when and how to respond, and emits control signals for what to say plus emotional tone, stance, facial expression, and gesture.
• Audio-Visual Generation - a streaming speech generator and a streaming video generator that turn those control signals into voice and video as they arrive.

The conversational engine makes its decisions at regular sub-second intervals instead of once per turn. That lets it read a pause for thought as thinking and keep waiting, and nod or back-channel while you are still speaking. Because it perceives video, it can also use your gaze, your expression, and anything you show the camera.

The generation side is built for streaming:

• Speech - an autoregressive diffusion transformer generates speech one latent chunk at a time, so it starts talking before the sentence is complete, and it can clone a voice from about 10 seconds of audio.
• Audio codec - Tavus's Tavec codec maps 48 kHz audio to 40 values per frame at 100 frames per second and streams audio packets as small as 10 ms.
• Video - a few-step autoregressive diffusion generator, distilled in three stages from a large many-step model, produces 720p video in 320 ms chunks in real time.
• Steering - streaming controls let the conversational engine push a gesture, a glance away, or a change of emotion into the very next chunk of video.

For testing, the architecture decides where failures show up. A cascade adds delay at every handoff and cannot react while the user is talking, while a full-duplex model can react at any moment, including the wrong one, so test turn-taking in both directions: when the agent should speak up and when it should stay quiet.

Tavus Griffin on NVIDIA's VideoFDB Benchmark

VideoFDB is NVIDIA's benchmark for full-duplex audio-visual conversation, built from 237 annotated clips of real two-person video calls that span 11 nonverbal conversational dynamics, such as turn-taking, backchannels, and gaze aversion. A language-model judge scores each response from 0 to 5 on two tracks: perception, for whether the agent reads the situation, and generation, for whether its own audio-visual output fits.

These are the overall scores on NVIDIA's leaderboard, where Griffin-Lite is the only entry listed on both tracks. The highest non-human score on each track is in bold:

On generation, Griffin-Lite finished 1.03 points ahead of the next system and 0.09 points below the human reference. On perception it led the strongest baseline by 0.29 points, the top score among the 15 entries evaluated.

NVIDIA also publishes the rubric scores behind each track, and they show where the remaining gap to humans sits. In the timing rows, the first figure is takeover-rate (TOR) alignment, how closely a system's choices about when to speak match the reference conversations, and the second is median latency:

The breakdown separates what Griffin already does well from what still lags:

• Emotional matching - dyadic affect is the one rubric where Griffin-Lite scores above the human reference, 4.40 against 4.14.
• Nonverbal cues - nonverbal cue appropriateness is its lowest generation score, 2.83 against 3.18 for humans, though still well ahead of the cascaded avatars.
• Response time - its median latency runs about a second behind the human reference on generation and 0.8 seconds behind on perception.
• Conversational flow - on perception, MiniCPM-o 4.5 in its audio-only configuration scores higher on flow, 3.76 against 3.67.

NVIDIA's own findings explain the size of Griffin's lead over cascades. Audio-driven avatars move only in response to the speech they produce, so they cannot add a nod or an expression while the user is talking, and the two cascaded systems in the benchmark had median latencies of 2.8 to 3.5 seconds.

Both tracks are scored by a model rather than by human raters, the same LLM as a judge approach many teams use to grade their own agents, so read small score differences with care.

How Fast Is Griffin's Video Generator?

Tavus also tested Griffin-Lite's video generator on its own, against four published streaming diffusion models that each turn speech and a reference image into a talking face. Tavus reports these results:

• Latency - true audio-to-video latency averaged 0.43 seconds on H100s, half that of the next fastest method, because the generator produces one latent at a time with no lookahead audio.
• Visual quality - first among the five models on DOVER and FID, the two standard video quality measures, and on THEval, a recently published framework built for talking heads.
• Lip sync - second on LSE-C, at 7.27. Tavus notes that LSE-C rewards pronounced mouth movement, including movement past the point where it looks natural, which THEval accounts for.

These comparisons come from Tavus rather than an independent evaluator, so weigh them below the VideoFDB results. In a conversation, the lower latency matters most when a person interrupts or a nod is due, because the face reacts sooner.

Did Tavus Griffin Pass a Video Turing Test?

Tavus recruited participants through an independent research platform and told them they would be matched with another participant for a one-minute call about what they were looking forward to this year. Only at the end of a survey were they asked whether it had crossed their mind that their partner might not be a real person. Every participant was then told the partner had been an AI.

Tavus ran the same protocol on its previous stack of Phoenix-4.5, Sparrow-2, and Raven-1:

Participants also rated Griffin-Lite on a 7-point scale:

• Seemed natural - 5.4.
• Seemed trustworthy - 5.6.
• Would enjoy talking again - 5.8, holding at 5.4 among participants who said it was an AI.
• Was really listening - 5.5.
• Conversation flowed naturally - 4.9, the lowest of the five.

Over half of the participants said the possibility never crossed their mind during the call, and those who did suspect tended to do so within the first 20 seconds.

Tavus designed and ran the study, the sample was small, and each call lasted one minute on a single light topic. Treat the headline result as a strong early signal, and note that flow, the rubric where an audio-only model beat Griffin-Lite on VideoFDB, also rated lowest with people.

Note: In Tavus's study, conversation flow rated lowest. TestMu AI Agent Testing scores conversation flow, question handling, response quality, and avatar presentation from 0 to 10 on a recorded video session, so you can see which one your agent is losing. Try TestMu AI free!

When Can You Use Tavus Griffin?

Not yet. Griffin-Lite is a research preview for select trusted testers, Tavus says it will not be available to customers for now, and a wider release of a more powerful model is to follow.

Tavus ties the release to safety work:

• Deception risk - Tavus says the properties that make Human Interaction Models natural also let them deceive a person into believing they are not talking to an AI.
• Disclosure features - it is building safe disclosure features and working with organizations tackling AI safety before release.
• Trusted-tester access - teams can request preview access through a form linked from the announcement.
• Available today - developers can build on Phoenix-4.5 for rendering, Raven-1 for perception, and Sparrow-2 for turn-taking, models Tavus says 150,000 developers and businesses already use.

The disclosure question reaches beyond Tavus. If your video agent could pass for a person, test that it identifies itself as an AI when asked directly and wherever your policies require it.

How to Turn Griffin's Benchmarks Into Video Agent Tests

A transcript cannot show whether an agent talked over a user, froze mid-sentence, or drifted out of lip-sync, and those are the areas Griffin's own scores point to. If the top-scoring system still falls short there, your agent probably does too, so I would map each finding to a test:

Agent Testing runs those checks on a live session. TestMu AI joins your agent's joinable web session URL as a simulated candidate with a real face and voice, holds an improvised conversation, and records it, with no SDK, no phone number, and no change to the agent.

• Personas and fan-out - one scenario expands across avatar faces, personas such as blunt, chatty, distracted, or impatient, test profiles, and up to 10 iterations, with up to 50 sessions per suite run.
• Success criteria - you write the pass conditions, the agent passes only if every criterion is met, and anything that cannot be verified from the recording counts as not met.
• Evidence - each run returns a downloadable recording, a turn-labeled transcript, a criteria table with Expected, Achieved, Evidence, and Confidence columns, and timestamped evidence.
• Fair scoring - the simulated candidate counts as test equipment, so its own glitches cannot lower your agent's result, and sessions that fail on the TestMu AI side are marked Inconclusive.

Video testing is enabled per organization, sessions run two at a time and last 30 to 600 seconds, and agents that exist only inside Zoom, Meet, or Teams are not supported yet. Point runs at a staging URL, because every session is a real conversation with whatever sits behind the link.

For the full method, see video agent testing and video simulation testing, and for a practitioner view, the Testμ 2026 session on testing video AI agents.

Getting Started With Tavus Griffin

If you need Griffin before its wider release, request trusted-tester access from Tavus, and keep building on Tavus's current models or your existing video stack until then. Write the success criteria and personas for your agent now, run them against today's version, and keep the results to rerun when you switch models; the guide to getting started with Agent Testing covers the setup.