A generated title card for the article 'Avatar 6 vs HeyGen Video 1', with a kicker reading 'AVATAR MODEL COMPARISON', the large title 'Avatar 6 vs HeyGen Video 1', the subtitle 'Same vendor, two different jobs', and three pill cards reading 'person in the frame', 'scene in the frame' and 'one credit account'.
Guides & Insights

Avatar 6 vs HeyGen Video 1: Same Vendor, Two Different Jobs

Author

Elias Hawthorne

Date Published

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

HeyGen shipped Avatar 6 on 17 November 2026, seven weeks after it shipped HeyGen Video 1 on 30 September. They sit next to each other in the same model menu, both accept a reference image, both output video with sound, and both are billed by the same account. Choosing between them is not a quality question — neither one is a worse version of the other. It is a question about what is in the frame. If a specific person has to appear and speak your words, only one of these can do it. If nobody in particular has to appear, only the other one can.

Start with the frame, not the spec sheet

The two models answer different prompts. Avatar 6 animates a person: you hand it a photo or a short video of someone and the speech audio they should perform, and it returns that person saying those words, moving the way the audio implies. HeyGen Video 1 generates a scene: you hand it a description, or an image, or a reference clip, and it returns five to thirty seconds of footage with its own dialogue, ambience and effects, with no presenter implied at all.

That distinction decides most real projects before any number is compared. A training module where a named instructor has to deliver a script is Avatar 6. A product ad where a coffee cup steams on a counter at dawn is HeyGen Video 1. A hybrid — an instructor presenting over generated b-roll — needs both, and HeyGen's own Video Agent exists to stitch them.

The spec contrast that actually differs

• What you send — Avatar 6 takes reference media plus audio; HeyGen Video 1 takes a text prompt, an image, or a reference clip, and writes its own audio.

• Output length — Avatar 6 runs as long as the audio you supply, with 25-to-28-minute continuous renders demonstrated; HeyGen Video 1 produces 5-to-30-second clips.

• Resolution — Avatar 6 offers 1080p and 2K; HeyGen Video 1 spans 480p to 2K, with 1080p and 2K at roughly three times the 768p rate.

• Identity — Avatar 6 holds one person's face, voice and mannerisms across the whole render; HeyGen Video 1 has no identity to hold, and consistency across clips is a prompting problem, not a model guarantee.

• Audio — Avatar 6 lip-syncs to audio you provide, so the voice is yours to choose and clone; HeyGen Video 1 generates its audio from the prompt, dialogue included, at roughly 2.5 words per second for timing estimates.

• Price — Avatar 6 is sold in subscription credits with no published per-second rate; HeyGen Video 1 has a published per-second API price, quoted in HeyGen's own material as $0.02 with a $0.03 figure appearing in its cost chart.

• Motion control — Avatar 6 takes a motion_prompt of up to 1,200 characters plus a two-level expressiveness switch; HeyGen Video 1 takes camera, lens and movement instructions inside the shot description.

A generated two-column scoreboard titled 'Avatar 6 vs HeyGen Video 1 — the scoreboard'. Left column 'Avatar 6': What you send reference media plus audio; Output length as long as the audio, 25-28 min shown; Resolution 1080p and 2K (up to 2880x1620); Identity one person held throughout; Audio lip-sync to audio you supply; Price credits, no per-second rate. Right column 'HeyGen Video 1': What you send a text prompt, image or reference clip; Output length 5-30 second clips; Resolution 480p to 2K; Identity no identity to hold; Audio audio generated from the prompt; Price $0.02/s published. The footer reads 'Video 1 rate per HeyGen model page. Avatar 6 per-second pricing not published.'

Two of those rows carry more weight than the rest. Length, because it is the difference between a clip and a deliverable. And price transparency, because it is the difference between a budget and a guess.

Where the price asymmetry bites

HeyGen Video 1 has a rate card. Its model page quotes $0.02 per second for API access, and the parameter table describes resolutions from 480p to 2K with 1080p and 2K costing about three times the 768p rate. A thirty-second 768p clip is therefore roughly sixty cents, and you can multiply that by your monthly volume without asking anyone.

Avatar 6 does not have one. The public pricing page sells Creator at $29/month for 600 credits and Pro at $49/month for 1,000 credits, and lists Avatar 6, Avatar 6 Turbo and Avatar 6 Pro as included capabilities — but no credit-to-second conversion is published, and the developer pricing page resolves to a logged-in app. The gap is not an oversight so much as a maturity difference: HeyGen Video 1 has been on the market since September and has settled into a per-second rate, while Avatar 6 is two days old and still being sold as a subscription feature.

For anyone modelling cost, that asymmetry inverts the usual advice. The newer, more impressive model is the one you cannot yet price per unit. Budget Avatar 6 in credits and validate the conversion with HeyGen before you scale, or pilot it on the free tier — three watermarked videos a month, one minute each — until a rate card appears.

What each one is actually good at

The number that predicts what you will see in a finished file is 65.1%, not the 84.4% on the banner. HeyGen's headline rescales a 42.2% side-by-side win rate — raters shown a genuine recording beside an Avatar 6 clip of the same speaker, same line, same audio — against a coin-flip baseline of 50. Shown a clip alone instead, 65.1% of Avatar 6 outputs passed as real against 78.9% of unedited recordings, and that figure fell to 56.5% once a HeyGen voice clone stood in for the person's own voice. The voice is carrying a large share of the persuasion, which is worth knowing when you decide who records it.

Outside HeyGen, exactly one measurement existed at the time of writing. New Measure published early blind-vote results from its Avatar Arena on 18 November, and Avatar 6 took first on both boards: 1,197 Elo for personal avatars, 1,237 for stock presenters. Read the limits before quoting it. Scripts were capped at 500 characters, 97% of completed comparisons used clips of 35 seconds or less, and three rival entries spoke through HeyGen's own cloned voice. Every one of those constraints describes the kind of clip HeyGen Video 1 produces, not the 20-minute render Avatar 6 is sold on.

A screenshot of HeyGen's Avatar 6 product page, showing the banner 'Introducing HeyGen Avatar 6, #1 on Avatar Arena', the heygen-avatar-6 model identifier, a 'Write a script' box containing a bracketed-emotion sample script at 195 of 320 characters, an 'Add motion' row with chips for Laughing, Whispering, Surprised, Angry, Excited and Custom, and the 'Choose a presenter' row.

HeyGen Video 1's claims run in a different direction. Its published evaluation is an internal set of Artificial Analysis Arena queries with 4,800 votes, scored on HeyGen's own normalised Elo, plus a 55.8% blind-preference win over another vendor's MiniMax H3 post-train on 650 votes with a 49-to-62% range — a range whose low end straddles a coin flip. HeyGen Video 1 does not appear on Artificial Analysis's public video boards. What it does have is a documented lineage: it is built on MiniMax H3 and post-trained by HeyGen, and its prompts are structured like a shot list, which tells you the model expects cinematographic direction rather than a sentence.

The reference-media rules are where the two diverge hardest

The consent gate is the asymmetry that never makes it into a pricing table. Feed Avatar 6 a video of a real person and the workspace needs Enterprise access plus an on-camera statement carrying a code minted for that session — a release form signed last month is not evidence, and nothing renders until HeyGen's match check passes. A photo reference or a pre-existing avatar look sidesteps the entire flow. HeyGen Video 1 has no equivalent gate, because a reference clip there steers style rather than reconstructing a named face; what you inherit instead are the ordinary licensing questions about the footage you uploaded.

HeyGen Video 1 has none of that. Its reference-to-video mode takes a reference clip to steer style and subject without the identity verification Avatar 6 demands, because it is not reconstructing a specific real person's face over time. If your use case involves consent-covered talent, Avatar 6's rules are a feature. If it involves a stylised character, they are simply not in your path.

Running both from one place

OrcaRouter does not serve either model. Both are HeyGen-hosted, consent-gated and billed in HeyGen credits, and no routing layer changes that. What a routing layer does change is the cost of the parts around them. The script an avatar reads, the shot list a video prompt is built from, the classifier that decides which of the two a request needs — those are ordinary model calls, and they are the ones that fail at 2am. Two hundred plus models behind one OpenAI-compatible key, at provider list price with zero markup, with automatic failover, means the renderer is the only vendor relationship you have to think about.

There is a second, quieter use. HeyGen Video 1 is built on MiniMax H3, which is a routable model at $0.08 per second for 768p and $0.13 for 2K. If you want to know whether HeyGen's post-training earned its premium, running the base model on the same prompt is a cheap experiment — and it is one you can do without a HeyGen account.

A screenshot of the OrcaRouter model page for MiniMax-H3 (minimax/minimax-h3), showing the model ID, the description that it generates 4-15 second videos at 768P or 2K with native stereo audio billed per second of output, the per-second price of $0.0800 at 768P and $0.13 at 2K, and OpenAI-compatible code samples.

Which one to reach for

Reach for Avatar 6 when the deliverable is a person. Explainer videos, localisation of an existing spokesperson, sales outreach where the sender's face matters, anything longer than a minute, anything where the audio is fixed and the mouth has to match it. The word-level edit mode matters more than it sounds: regenerating one sentence in a twenty-minute video while leaving 48.3 dB of the untouched footage intact is the difference between a five-minute fix and a full re-render.

Reach for HeyGen Video 1 when the deliverable is a scene. Advertising, b-roll, stylised characters, anything where a shot list is the natural way to describe what you want, and anything short enough that thirty seconds is not a constraint. The per-second price and the published duration limits make it the easier of the two to put in a budget.

Reach for both when you are producing presenter-led video with cutaways, which is most corporate video. HeyGen's Video Agent composes them, and the split is clean: the presenter is Avatar 6, the cutaways are HeyGen Video 1, and neither model has to do the other's job badly.

What is still unmeasured

Nobody outside HeyGen has tested Avatar 6 beyond 35 seconds. The arena capped out there, and every long-form claim — 25 to 28 minutes, continuous, one generation — comes from HeyGen's own demonstrations. Whether identity drifts at minute twelve, whether the gesture vocabulary starts repeating, whether a 2K render of that length is economically sane: all open. HeyGen Video 1's open question is the opposite one, and older: whether a thirty-second clip ceiling is a technical limit or a product decision, and whether the price holds once the launch window closes.

Buy the one that matches your frame. Do not buy either on the strength of a preference percentage — including HeyGen's, and including the arena's.