Tavus launches Griffin, real-time video conversation model that fools 48% of testers
On October 2, digital human company Tavus released Griffin, a real-time video conversation model positioned as the first "Human Interaction Model," unifying real-time perception, dialogue, and expressive video generation into an end-to-end video-to-video system, and claiming to be the first model to pass a "Video Turing Test." It is currently available only as a limited research preview, and the Turing-test claim rests on official and early-tester accounts without independent third-party verification.
Confirmed
- Tavus officially released Griffin for real-time conversation scenarios; @testingcatalog, @rohanpaulai, @lmoroney and @APPSO relayed the official announcement.
- Official test data: after one-minute video calls with the preview Griffin-Lite, 48% of 54 participants believed Griffin was human, versus under 3% for prior systems.
- Griffin ranked first in NVIDIA's full-duplex AI video evaluation / independent panel review (per official and early-tester accounts).
- Developer @TejasKumar found Griffin to be a "never-stopping conversation model": it listens and watches continuously, deciding each moment whether to speak, react, or wait—breaking the turn-based interaction paradigm—with speech starting before video generation completes and only 0.43s audio-video latency.
- @jaydeepkarale relayed that Griffin is a limited research preview, with the team sharing progress live ahead of a safe public release.
- @paulcal relayed a "Simon Says" demo where two humans, Brittney and Mars, tried to trick Griffin, which had to simultaneously hear the command, observe gestures, and perform the action—showing synchronized listening, watching, and acting.
- Early tester Jason Kneen shared hands-on impressions and relayed rumors that Griffin is not publicly released yet because it is too capable.
- @Distinct-Question-16 and @DrSingularity posted demo videos showing near-indistinguishable real-time avatars; @DrSingularity also imagined interacting with it on 65-inch-plus screens.
- @alexisgallagher relayed kwindla's view that voice agents in real deployments already passed the Turing test—users often forget they are talking to an agent even when told upfront.
Unconfirmed
- The "passed the Video Turing Test" claim is promotional, without detailed methodology, and awaits third-party validation.
- Some posts (@Distinct-Question-16, @DrSingularity) cite a 44% misidentification rate, slightly diverging from the official 48%.
- @markk cautioned there is no official source or third-party verification; @Puzzleheaded-King584's Reddit post did not identify the model.
- "Withheld from public release for being too capable" remains an unconfirmed rumor; the official stance is gradual rollout via limited research preview.
Why it matters
- The 48% human-mistaken rate is an order-of-magnitude jump from under 3%, which—if independently verified—would significantly advance the usability frontier of real-time digital human interaction.
- The 0.43s latency and never-stopping architecture shift interaction from turn-taking to continuous perception and real-time reaction, and the Simon Says demo shows genuine multimodal coordination rather than mere visual realism.
- Griffin remains a limited research preview; its capability boundaries and commercialization path are still uncertain and worth continued follow-up.
2026-10-02 ~ 2026-10-03 · 17 related posts
Primary sources
- Tavus' Griffin passes video Turing test: 48% of live viewers mistook it for a human — EXM7777 · 2026-10-02
- Tavus's Griffin claimed to pass face-to-face Turing test, deemed too powerful to release — jasonkneen · 2026-10-02
- Viral Claim: Real-Time Conversational Video AI 'Griffin' Passes the Turing Test (Unverified) — mark_k · 2026-10-02
- [source] Tavus unveils Griffin, first model to pass real-time video Turing test with 48% — testingcatalog · 2026-10-02
- Griffin video avatar fools 44% of viewers on NVIDIA full-duplex benchmark — Distinct-Question-16 · 2026-10-02
- Tavus's new realtime video model flirts with passing the conversational video Turing test — alexisgallagher · 2026-10-02
- Tavus digital human Griffin plays Simon Says, generating full frames in real time — paul_cal · 2026-10-02
- First model to pass the 'Video Turing Test'? Half of people thought it was human — Puzzleheaded-King584 · 2026-10-02
- [source] Tavus's Griffin starts talking before generation finishes, face reacts in 0.43s — TejasKumar_ · 2026-10-03
- Tavus claims Griffin passes the video Turing test as 48% mistake it for a human — APPSO · 2026-10-03
- Tavus previews Griffin, a Human Interaction Model that sees, hears and speaks like a person — _jaydeepkarale · 2026-10-03
6 near-duplicate retellings: rohanpaul_ai · Dr_Singularity · Dr_Singularity · Dr_Singularity · Puzzleheaded-King584 · lmoroney