A generated title card for the article 'Avatar 6 vs Tavus Griffin', with a kicker reading 'AVATAR MODEL COMPARISON', the large title 'Avatar 6 vs Tavus Griffin', the subtitle 'One is buyable today, the other is a preview', and three pill cards reading 'public API, credits', 'application only' and 'two different tests'.
Guides & Insights

Avatar 6 vs Tavus Griffin: One Is Buyable Today, the Other Is a Preview

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Two avatar models, six weeks apart, both shipping within 2026, both claiming a Turing-test result. Avatar 6 went live on 17 November 2026 with an API, a free tier and two model identifiers. Tavus Griffin was announced on 1 October 2026 as the world's first Human Interaction Model and shipped as Griffin-Lite, a research preview available to a select group of early testers through a request form. That is the whole comparison in two sentences, and everything below is the detail that decides whether it matters to you.

Because it does matter, and not in the direction people expect. Griffin is doing something Avatar 6 is not attempting, and the gap is not incremental. It is the difference between a renderer that performs a script and a system that decides when to speak. That the more ambitious one is also the one you cannot buy is the tension this page has to hold.

What "preview" means at Tavus, specifically

Griffin is not on the Tavus platform. There is no model identifier to pass to an endpoint, no price, no concurrency limit, and no published date for general availability. Tavus's product menu lists it as "Griffin (Preview)" alongside Phoenix-4.5, Raven-2 and Sparrow-2. Access is a request form and a selection process. The company has said a wider release of a more powerful model will follow, without saying when.

Avatar 6, by contrast, has documentation you can read this afternoon. Two identifiers — heygen-avatar-6 and heygen-avatar-6-turbo — on two endpoints: POST /v3/models/avatars for reference-media-plus-audio, and POST /v3/videos with engine.type set for existing avatar looks. A 202 returns a video_id, you poll, you get a signed URL. The free tier runs three watermarked videos a month at up to a minute each.

So the first honest sentence of any Avatar 6 vs Griffin comparison is that one of them is a purchase and one is an application. Everything else is a comparison of published claims.

The two Turing tests are not the same test

Both vendors lead with a number near 50%, and neither number means what a reader will assume.

Tavus: 48% of people who talked with Griffin for a minute believed they had spoken with a real human. The company notes that previous systems, including its own stack powered by Phoenix 4.5, Sparrow-2 and Raven-2, had a maximum pass rate of 2%. That is a live, open-ended, back-and-forth video call, judged by the person having it. The comparison is against Tavus's own earlier systems, and the sample is not published in the launch material.

HeyGen's figure comes from a file comparison, not a conversation. Raters saw a genuine recording beside an Avatar 6 clip of the same speaker delivering the same line over the same audio and picked Avatar 6 42.2% of the time; the 84.4% on the banner is that rate divided by 50 and doubled, which measures distance travelled toward perfect confusion, not how often anyone was fooled. Shown a clip alone, with nothing to compare it to, 65.1% of Avatar 6 outputs were judged real against 78.9% of unedited recordings.

Put them side by side and the asymmetry is the point. HeyGen asks whether a static performance survives comparison with a known reference. Tavus asks whether a live back-and-forth survives a minute of improvisation — interruptions, subject changes, silences that need filling. A model can be excellent at the first and useless at the second, and nothing in either vendor's numbers tells you where the other one would land.

Architecture, and why it changes the product

• Input — Avatar 6 takes a reference image or reference video plus speech audio you supply. Griffin takes a live audio and video stream from the person talking to it.

• Decision-making — Avatar 6 performs what it is given. Griffin runs a conversational model that decides when to speak, hold or yield, at sub-second intervals rather than once per turn.

• Output — Avatar 6 returns a video file at up to 2880×1620, as long as the audio. Griffin returns a live video-to-video stream in real time.

• Duration — Avatar 6: 25 to 28 minutes demonstrated as single continuous generations. Griffin: bounded by the call, not by a render.

• Latency — Avatar 6: not published; the independent arena measured a typical 175.6 seconds per render, with one in ten over nine and a half minutes. Griffin: real-time by construction, with Tavus reporting it 37% ahead of the next best AI at reacting in the moment on NVIDIA's VideoFDB.

• Access — Avatar 6: public API, credits. Griffin: research preview, application only.

• Price — Avatar 6: subscription credits, no per-second rate published. Griffin: none.

Look at the first two rows together. Avatar 6's input is finished speech, which means the hard part — knowing what to say and when — happened somewhere upstream, in whatever wrote the script. Griffin's input is a person, which means the hard part is inside the model. That is why one is a renderer and the other is being sold as a new class.

A generated two-column scoreboard titled 'Avatar 6 vs Tavus Griffin — the scoreboard'. Left column 'Avatar 6': Turing-test claim 42.2% picked as real; Input reference media plus audio; Output video file up to 2880x1620; Latency 175.6 s typical render; Access public API, credits; Price subscription credits. Right column 'Tavus Griffin': Turing-test claim 48% believed human after a minute; Input live audio and video stream; Output live video-to-video stream; Latency real-time by construction; Access research preview, application only; Price none published. The footer reads 'Both headline figures are vendor-reported. Griffin has no public likeness benchmark.'

Griffin's engineering reflects it. Tavec v1.7, its audio codec, maps 48 kHz audio into a continuous latent of 40 values per frame at 100 frames per second with no codebooks and a fully causal streaming decoder, reporting 2,193× real-time decoding and a 4.242 ViSQOL MOS-LQO at 28.42 million decoder parameters, measured on an H100. Voice cloning works from about ten seconds of audio. The video side generates every pixel of the scene in real time from one reference image — not just the face and hands but the chair the person sits in and the shadows they cast. None of that is applicable to Avatar 6, which has a render budget measured in minutes and can afford to be expensive.

What the independent evidence covers, and what it does not

Neither model has been through a public head-to-head. That is worth stating without hedging.

The one figure here HeyGen did not produce is a leaderboard position. New Measure's blind arena ranked Avatar 6 first on both boards on 18 November — 1,197 Elo on personal avatars, 1,237 on stock presenters, Creatify Aurora 2.0 second on each — and was candid about the ceiling: short scripts, clips of 35 seconds or less in 97% of comparisons, three entrants speaking through HeyGen's cloned voice. Griffin has no equivalent entry, because a preview behind a request form cannot be ranked.

A screenshot of the New Measure article 'Which AI avatar do creators prefer?', published 18 November 2026, showing the finding that HeyGen Avatar 6 was preferred most often on both boards with Creatify Aurora 2.0 second, and the beginning of the blind-vote methodology.

Griffin has NVIDIA's VideoFDB, where Tavus claims first place on an independent test of face-to-face AI and a 37% lead over the next best system at reacting in the moment. That is a real third-party benchmark rather than a vendor preference study, and it is the strongest single piece of evidence in this comparison. It is also a benchmark of a capability Avatar 6 does not have, so it does not settle anything about likeness.

What neither has is a comparison against the other, under one protocol, run by someone with no product to sell.

Where the two diverge hardest is what you must prove before anything renders. Avatar 6 will not animate a real person from video until the workspace holds Enterprise access and a consent recording made on camera, carrying a one-time code issued for that session; HeyGen then compares the consenting face against the source footage. Without video-reference access the request fails with a 422 realistic_human_reference_not_allowed rather than falling back to something cheaper.

Griffin's constraints are different in kind because its input is a live person rather than a stored likeness. The consent problem it raises is not "may I clone this face" but "what does this system retain from a call" — and Tavus's launch material addresses safety and limitations without publishing the kind of per-render consent gate HeyGen has built. For a deployment involving real people on camera, that distinction matters more than any benchmark.

Where OrcaRouter comes in

OrcaRouter sells neither of these. One is a HeyGen render job behind a consent gate and a credit balance; the other has no endpoint to call. We are not going to imply otherwise in order to make a routing pitch land.

What we do serve is the text between the two ends of either pipeline. Avatar 6 opens with a script someone has to write, review and localise; Griffin's conversational layer has to hand off the tasks it cannot finish mid-call, and Tavus routes those out to other models by design. Those are commodity calls — summarise, translate, classify, retrieve — and they are the ones that benefit from a single key across 200+ models at provider list price with zero markup, and from failover that turns a provider incident into a retry instead of an outage.

A screenshot of the OrcaRouter Models page, showing the filter sidebar and the header reading 226 models, with the All 226, Cyber 11, Text 190 and Image 12 tabs.

The decision, as it stands on 18 November 2026

For a presenter delivering your script at volume, with a documented API and a production path this month, Avatar 6 is the only eligible candidate of the two. It tops the one independent board that has scored the category, and it comes with three items to carry into a pilot: no per-second price on any public rate card, no outside test longer than 35 seconds, and renders that take minutes rather than seconds.

If you need a system that can hold a live conversation with a face attached, Griffin is the more advanced work and the less available one. Apply for the preview if the use case is real. Do not put it on a roadmap with a date, because Tavus has not given one.

The thing to watch is not a benchmark. It is whether Tavus ships Griffin generally, and at what price — because a real-time interaction model with an endpoint and a rate card would not compete with Avatar 6 on likeness. It would make the category Avatar 6 belongs to look like the previous generation.