Fix One Word, Re-Render 34% of the Frame: HeyGen's Avatar 6 Keeps a Talking Head Going for 28 Minutes — but Its Trophy Comes From a 35-Second Arena
TL;DR
HeyGen released Avatar 6 on November 17. It is a talking-head model: give it a photo and an audio track, and the person in the photo delivers your script with facial expression and hand gestures.
That pitch is two years old. What changed is three specific annoyances.
First, gestures stop looping. The giveaway in most AI presenter video is not the face, it is the hands — the same wave, the same nod, the same open palm on repeat, and the illusion collapses within a minute. Avatar 6 reads the script and attaches movement to specific words: a thumbs-up on "hello," a finger pointed sideways on "the counter on the left."
Second, it can run for half an hour. HeyGen's demos include four clips of 25 to 28 minutes, each generated in one pass from a single audio track, with no stitching and no loops. Length follows the recording instead of being capped at five or ten seconds.
Third, fixing a line no longer means re-shooting the take. This is the practical one. Misspeak one word and the old workflow regenerates everything — face, gestures, background all drift. Avatar 6 re-renders only the edited span and leaves the rest untouched.
The trade-offs are just as concrete: it is slow, and its headline "number one" ranking deserves a closer look.
How it works
Gestures tied to words, not timestamps
Avatar 6 accepts a motion prompt: you attach an action to a phrase, and the model places it on that phrase while keeping the speaker's own movement habits in between.
HeyGen's examples are concrete. A wave on "Welcome." A thumbs-up on "grandparents." A hand on the heart on "Thanks for trusting us." You can also specify only mood or expression and leave the gesture to the model.
The interesting part is not that you can request a gesture — plenty of tools allow that — but that the request is bound to a word rather than to a timecode. Edit the sentence and the gesture follows the new line automatically.
One continuous 25-to-28-minute take
The technical report shows four full videos of 26, 26, 25 and 28 minutes, each a single continuous generation from one audio track with no looping.
The number itself is not shocking; you can stitch short clips into 28 minutes. The difference shows at the seams — a lighting jump, a posture reset, a fold of clothing that changes. Continuous generation has no seams, but it demands that the model not drift over half an hour, which is a different problem.
The same framework also handles two people in one frame: write both sides of the script and it generates a host-and-guest conversation in one go. The demo is a two-minute podcast that cuts between a wide two-shot and closer views of each speaker.
Editing one word touches 34.4% of the frames
HeyGen ran 20 edit tests — change a word, add a gesture — and counted how many frames actually changed:
- Avatar 6: 34.4% on average
- Seedance 2.5: 99.7%
- MiniMax H3-max: 100%
- Wan 3.0: 100%
- HappyHorse Edit: 100%
For the untouched region, HeyGen reports PSNR against the original: 48.3 dB for Avatar 6, versus 29.3 dB for the next best, Seedance 2.5. Higher means closer to the source footage.
In plain terms: the other models re-shoot the scene from the new script; Avatar 6 cuts into the existing footage. For a finished presenter video that is one word short of done, that is the difference between an hour of re-recording and a few minutes of waiting.
Reading the numbers
The three most quoted figures are 84.4%, 72.4% and first place on Avatar Arena. Each needs a qualifier.
84.4% does not mean 84% of viewers think it is real
The precise result: across 270 side-by-side answers where raters picked which of two clips was the genuine recording, 42.2% chose the Avatar 6 version. Since a two-way choice has a 50% baseline, dividing 42.2% by 50% yields 84.4% — that is how the "indistinguishability" metric is defined.
The test involved 20 external raters and 76 people on camera. It is not a mass blind test, and the sample is modest.
Shown alone, it still gets caught
In a separate test, raters saw one clip and judged whether it was real:
- Avatar 6 with the person's own voice: 65.1% judged real
- Genuine recording: 78.9% judged real
- Avatar 6 with a HeyGen stock voice: 56.5% judged real
Swap in a synthetic voice and the detection rate rises noticeably. What sells the illusion is not picture alone.
Against general video models, the lead narrows
From the person's own footage, with external raters expressing a preference:
- Versus Seedance 2.5: 72.4% favor Avatar 6
- Versus Wan 3.0: 74.1%
- Versus Kling 4.0: 54.9%
- Versus Wan 3.0 starting from a single photo: 50.1%, which HeyGen itself labels a tie
From a photograph, it is roughly level with Wan 3.0. That is not a knock; it just means the specialist advantage shows up mainly when you have real footage and need a long, coherent talking performance.
The top ranking comes from short clips
Avatar Arena, run by New Measure, placed Avatar 6 first in both categories with Elo scores of 1,197 for personal avatars and 1,237 for stock presenters, against a category average of 1,000. Votes were counted from October 27 to November 16.
New Measure flags its own limits: participants opted in, the median script was 62 words, and in 97% of completed comparisons both clips were 35 seconds or shorter. That does not establish a lead on long video or in languages with thin test coverage.
It is slower than the competition
One more line in the same data is easy to miss: HeyGen personal-avatar videos took about three minutes on average, versus about one minute for Runway's GWM-1 Avatars, and one in ten HeyGen personal-avatar jobs ran past nine and a half minutes. These timings exclude avatar setup and speech preparation.
So "best looking" and "fastest" are two different products this round.
How to use it
Where: web, mobile and API. The API offers two paths — the Model API at POST /v3/models/avatars, which takes a reference image or video plus speech audio, and the Application API at POST /v3/videos, which takes a pre-built avatar look plus a script and voice. Model names are heygen-avatar-6 and heygen-avatar-6-turbo.
Resolution: 1080p by default; the Model API can request 2k. The "native 2K" in the technical report means 2880×1620 generated per frame, not an upscale.
Building an avatar: HeyGen says 30 seconds of footage is enough, with optional extra material or imports from social media. Photo references are open to all API workspaces; realistic video references are enterprise-only.
Consent is mandatory: a footage-based avatar cannot render until the person records an on-camera statement containing a session-specific one-time code, and HeyGen verifies that the consent speaker matches the person in the training footage.
One default that will bite you: in the Application API, omitting the engine selection continues to use Avatar IV, not Avatar 6. Avatar 6 on that endpoint also supports single-avatar requests only — no raw-image jobs, studio scenes, batch creation or scene editing.
Who it's for
- Teams producing a steady stream of presenter video — courses, product updates, sales outreach — where each clip matters less than the volume.
- Anyone holding a finished recording that needs one word fixed or one gesture added. Re-rendering only the edit is the feature that pays for itself.
- Long single-speaker talks that must hold identity for 25 minutes or more. Few products currently do continuous generation at that length.
- Podcast and interview formats with two speakers in frame, generated in one pass instead of composited in post.
Who it's not for
- Commercials that will be inspected frame by frame. The 65.1% versus 78.9% gap says it is still detectable.
- Multilingual rollouts into languages without solid test coverage. The published evaluations are essentially English.
- High-throughput pipelines that need output in under a minute. At roughly three minutes per clip, with a long tail past nine, plan capacity accordingly.
- Teams that need open weights and self-hosting. Avatar 6 is a closed commercial model.
Caveats
Do not read 84.4% as passing a Turing test. It means the share of raters picking the generated clip as real doubles the coin-flip baseline. Shown alone, 65.1% call it real, which leaves more than a third who do not.
Long video has no independent verification. The 25-to-28-minute films come from HeyGen. Third-party evaluation covered clips of 35 seconds or less in 97% of cases.
The leaderboard has a sample problem. Opt-in voters, a 62-word median script, sub-35-second clips — the setup favors short presenter spots and says nothing about half-hour talks.
Subscription price is not the cost. Credits scale with model, duration and complexity, and HeyGen publishes no single credits-per-minute table — only the hard limits: three videos a month on Free with a one-minute cap, and a 30-minute cap per video on paid plans.
2K and 4K are different claims. The report describes native 2K at 2880×1620; 4K export is a Pro-tier feature produced by super-resolution, not by the base model.
Cost and time
- Free: $0, three watermarked videos a month, up to one minute each, Avatar 6 access included.
- Creator: $29/month, 600 credits, Avatar 6 Turbo, 1080p export.
- Pro: $49/month, 1,000 credits, Avatar 6 Pro, 4K export, up to 30 minutes per video.
- Higher Pro tiers scale credits from 1,000 at $49 up to 100,000 at $4,300.
- Time: about three minutes per personal-avatar video on average, with one in ten past nine and a half minutes, excluding setup and speech preparation.
For a dozen presenter clips a week, Creator is likely enough. For a single 20-minute course video, check your credit burn first — there is no published per-minute rate.
First steps
- Decide whether your need is short or long. If you are making 30-second social clips, the long-form strength is wasted and cheaper tools will do.
- Take an existing take you had to abandon and fix one word. That is the capability with no real equivalent: check whether the surrounding frames hold still and the lighting stays put.
- Do not buy on the leaderboard. Test with your own script, your own face and your actual language, because the arena ran on English, a 62-word median script and clips under 35 seconds.
- Before wiring up the API, set the engine field explicitly. Leave it out and you get Avatar IV; video references require enterprise access and pass through the one-time-code consent check.
- Put render time in your schedule. Three minutes is the average, and the tail runs past nine.
The short version: Avatar 6 attacks the hardest part of making a synthetic presenter convincing — gesture and long-range consistency — plus the rework cost of fixing one word in a finished take. Its weaknesses are equally clear: it is slow, and outside short clips its lead has not been independently verified.
This article is compiled from HeyGen's technical report, product page, API documentation and pricing page, together with coverage from Unite.AI and PR Newswire. All percentages and durations come from those sources; no independent testing was performed.
Sources
- https://www.heygen.com/research/avatar-6-model
- https://www.heygen.com/models/heygen-avatar-6
- https://developers.heygen.com/docs/models/avatar-6
- https://www.heygen.com/pricing
- https://www.unite.ai/heygen-launches-avatar-6-with-script-aware-gestures-and-2k-video/
- https://finance.yahoo.com/technology/ai/articles/heygen-introduces-avatar-6-audio-170700762.html