HeyGen released Avatar 6 on 17 November, an audio-driven avatar model that builds a talking double of a person from a single photograph, with an optional short recording as an identity reference, and renders every frame natively at 2880 by 1620. The model ships in the web app, the mobile app, the HeyGen API and through the company MCP server.
The headline specification is duration. HeyGen published demonstrations of 25 to 28 minutes per video, each a single continuous generation from one audio track with no loops, so the running time simply follows the speech. That is an order of magnitude beyond the short clips that most avatar products are validated on, and it is the difference between a tool for social cuts and one that can carry a training module or a full product walkthrough.
The second change is motion. HeyGen says Avatar 6 uses the meaning of the script to decide where pointing, counting or directional movement belongs, so a written instruction such as a wave at a greeting or a hand on the chest at a reassurance is timed to the words that trigger it while the subject keeps their own habitual movement in between. Users can also specify mood and facial expression, or dial expressiveness down for calmer delivery. Two avatars can share one frame, which HeyGen demonstrated with a two-minute host-and-guest podcast generated from one script.
The editing model may be the more practical advance. Its recorded-video editor regenerates only the span that was changed, replacing a word and synthesising the new speech in the speakers own voice while leaving the rest of the take intact. In a 20-edit comparison published with the model, an average of 34.4 percent of frames changed, against 99.7 to 100 percent for MiniMax H3-max, Wan 3.0, Seedance 2.5 and HappyHorse Edit. For anyone who has re-recorded a 20-minute explainer because a single figure was wrong, that is the feature that matters.
The realism claim needs care. HeyGen reports 84.4 percent indistinguishability, which is not the share of viewers who were fooled: in side-by-side tests with 20 external raters and 76 participants, 42.2 percent of 270 decisive answers picked the generated clip as the real one, and the score doubles that selection rate because 50 percent would be chance. In single-clip tests a genuine recording was judged real 78.9 percent of the time, against 65.1 percent for Avatar 6 using the persons own voice and 56.5 percent with HeyGen Voice. All of these figures come from HeyGen research and have not been independently verified.
There is independent evidence of a sort. The evaluator New Measure ran a blind arena between 27 October and 16 November, before the public launch, and ranked Avatar 6 first for both personal avatars and stock presenters, with Elo-scale preference scores of 1,197 and 1,237. New Measure itself cautions against reading that as a long-form result: 97 percent of completed comparisons involved clips of 35 seconds or less, scripts were capped at 500 characters, voters were self-selected, and HeyGen took about three minutes per comparison against faster rivals.
Access is credit based. Free users get three watermarked videos a month of up to one minute; Creator is $29 a month for 600 credits, 1080p exports and Avatar 6 Turbo; Pro is $49 for 1,000 credits and Avatar 6 Pro, with Creator and Pro videos capped at 30 minutes and motion control limited to three uses a month on Free and 20 on Creator. Avatars of real people require live identity verification and on-camera consent, and HeyGen accepts removal requests.
What HeyGen is really selling is not a better face. It is continuity: a take that survives 28 minutes, an identity that holds across an edit, and a gesture plan that follows meaning instead of repeating the same shrug. Those are production properties, and they move avatar video out of the clip farm. The open questions are the familiar ones for synthetic media, namely who verifies consent at scale and how a viewer is supposed to tell a 28-minute generated presenter from a recorded one when the labels live outside the file.
Comments (0)
Log in to join the discussion
Log InNo comments yet