MarketingNew Release8 min readPublished

HeyGen Avatar 6 and Turbo: 2K Video API and Real Limits

A recorded presenter, a finished audio track and a saved avatar look lead to different production choices. Compare the direct render rates, the actual frame sizes and the evidence behind HeyGen’s realism claim.

DA
Digital Applied
PublishedNovember 19, 2026
Read time8 min

A production brief asks for six 45-second presenter videos in 2K. Through HeyGen’s Model API, a landscape reference produces 2880 × 1620 pixels. Through its Application API, a landscape avatar look produces 2560 × 1440. At direct list prices, the six first takes cost $23.63 on Turbo or $39.15 on the full model. Before anyone books a shoot or approves a budget, the producer needs to know which frame and which presenter the client agreed to.

HeyGen released Avatar 6 and its Turbo variant on November 17. Both animate a speaking presenter using an image or eligible video reference, or a look already available to the workspace. The practical use is a recurring address to camera whose words change between markets or releases. When the value of the shot is seeing a real person demonstrate an object, react to another person or attest to an event they witnessed, a recording of that event still serves a different purpose.

Key takeaways
  1. 01
    Choose the route before writing the jobThe Model API takes speech audio and reference media. The Application API accepts an eligible avatar look and either a script with a voice or supplied audio. Avatar 6 currently supports single-avatar requests on that route.
  2. 02
    Budget the currency your integration usesDirect API rates run from $0.05 to $0.145 per output second. The web product lists subscription credits per minute. An API key and an OAuth-backed creative tool can draw from different balances.
  3. 03
    Read the raw choices behind the realism headlineTwenty external raters took part in HeyGen’s test. Of 270 decisive paired choices, 114 selected the generated clip as the real one. HeyGen doubles that 42.2% share to produce 84.4%.
  4. 04
    Approve the reference and the intended useA photo works across workspaces with the subject’s permission. A realistic video reference on the Model API requires enterprise access. Creating a digital twin normally requires a separate consent check.

01 — InputsWhich assets do you already have?

The engine IDs are heygen-avatar-6 and heygen-avatar-6-turbo. HeyGen calls the full model Pro in the web product. The direct API has two ways to request either engine. HeyGen’s developer guide Avatar 6, checked November 19, specifies the input, the model field and the status route for each. Those differences matter when adapting a job that already generates video from a template.

Model API
Reference media and finished speech

POST /v3/models/avatars takes speech audio plus one image, one video, or both. If using both, provide a matching image of the same subject. A reference video must be at least two seconds. Set the model field explicitly. No existing avatar or voice setup is needed.

Use it for a recorded narration paired with an approved portrait. Poll GET /v3/models/videos/{video_id}.
Application API
A reusable avatar look

POST /v3/videos takes type avatar, a compatible avatar look ID and one speech source. Use a script with a voice ID, an audio URL or an uploaded audio asset. Set engine.type explicitly. Omitting it selects Avatar IV. Check supported_api_engines on the look.

Use it when production already owns an approved digital twin or uses a stock look. Poll GET /v3/videos/{video_id}.

The Application API’s Avatar 6 support does not extend to raw-image requests, studio scenes, batch creation or scene editing. For six localized versions, treat each presenter clip as its own request, then assemble the title, product footage, captions and closing card in the edit. The web studio can offer a different set of tools. A capability advertised for an editor or demonstrated in a research video does not establish support for the corresponding API request. Building a creative workflow around a video API

OpenRouter also lists both engines under heygen/heygen-avatar-6 and heygen/heygen-avatar-6-turbo. Those names identify the routed models. Integrations calling HeyGen directly should use its unprefixed IDs and choose the matching request shape. OpenRouter model listing · OpenRouter Turbo listing

02 — Delivery frameWhat does 2K actually return?

Both APIs default to 1080p and accept 2k. For an ordinary 16:9 reference at 1080p, either route gives you 1920 × 1080. At 2K the Model API uses a 1620-pixel short edge and keeps the reference shape unless you supply explicit output dimensions. Its 16:9 frame is 2880 × 1620, and portrait is 1620 × 2880. Optional width and height resize the finished output and must be supplied together as even pixel values.

The Application API uses the requested aspect ratio. Its 2K landscape frame is 2560 × 1440, portrait is 1440 × 2560, square is 1440 × 1440 and 4:5 is 1440 × 1800. A portrait video for a store listing and a square testimonial can therefore come from the same approved look with predictable export dimensions. Neither API offers a 4K Avatar 6 render in the documented resolution field. Avoid promising 4K detail because a different studio plan advertises a 4K export. When to crop, reframe or regenerate a video

Both routes accept a motion prompt of up to 1,200 characters. A direction such as “Keep your hands below the caption line and look at the camera on the final sentence” gives the reviewer something specific to inspect. For a quieter performance, the expressiveness setting accepts less or default. It lives at the top level on the Model API and inside engine on the Application API. A specific direction in the motion prompt takes precedence where the controls overlap.

03 — Rate cardAPI dollars and studio credits

The direct USD prices below are the four Avatar 6 rates in HeyGen Platform’s Self-Serve Pricing catalogue on November 19. Multiplying each per-second rate by 60 gives the cost of one generated minute. The subscription figures come from HeyGen’s help article How to use Avatar 6 (VI) in HeyGen. They apply to both photo avatars and digital twins in the web product, subject to the plan’s available modes and resolutions.

Billing basisAvatar 6 TurboAvatar 6 (Pro)
Direct API · 1080p$0.05/sec · $3.00/min$0.10/sec · $6.00/min
Direct API · 2K$0.0875/sec · $5.25/min$0.145/sec · $8.70/min
Web app · 720p or 1080p20 credits/min40 credits/min
Web app · 2K35 credits/min58 credits/min

The API-key route uses the API account’s balance or contracted rates. HeyGen’s developer documentation says its MCP integration uses OAuth and the user’s web subscription credits. A team that has tried a render from a chat application can therefore see a different charge when it moves the workflow to an API key. The presence of a 720p row on the web rate card also does not add a 720p option to either Avatar 6 API.

04 — Worked estimateWhat does a retake cost?

For the six localized 45-second videos in the brief, the first pass totals 270 output seconds. Turbo at 1080p costs $13.50. Full Avatar 6 at 1080p costs $27. At 2K the unrounded totals are $23.625 for Turbo and $39.15 for full Avatar 6. The corresponding 2K web usage would be 157.5 or 261 credits at the published per-minute rates before any account-specific rounding. Treat these as alternative ways to budget the job.

Suppose two translations come back for pronunciation changes. Rendering both complete 45-second clips again takes the total to 360 seconds. The direct 2K budget becomes $31.50 on Turbo or $52.20 on Avatar 6. Budgeting from the final program length alone would have left out the ninety seconds you needed to approve it. The estimate covers avatar renders. Voice generation, translation, editorial review and compositing may be separate costs.

A short comparison before rendering all six can settle whether full Avatar 6 is worth the difference for this client. Generate the same approved 15-second excerpt on both engines at the intended 2K setting. Together those two samples cost about $3.49 in direct render charges. Ask the presenter and editor to judge a tricky name, the pause before a key number, hand placement and framing at the actual display size. Lock the speech audio before committing the rest of the budget.

05 — Research evidenceWhat does 84.4% indistinguishability mean?

HeyGen Research’s November 17 technical report, Avatar 6: Talking Avatars That Are More Like You, describes a vendor-run test with 20 external raters. For the paired task, the team used 76 people’s footage. It kept one part as reference and held another back. The generated clip and the held-out real recording delivered the same words over the same audio. Raters saw the pair and indicated which was real, or that they could not tell. The evaluated segments with original speech ranged from 11.6 to 44.9 seconds.

There were 304 answers. Thirty-four were undecided, leaving 270 decisive choices. Raters chose the generated clip 114 times, or 42.2% of those choices, and the real recording 156 times, or 57.8%. HeyGen takes the generated clip’s share and divides it by the 50% expected from a random choice. Equivalently, it doubles 42.2% to obtain 84.4% on its scale. Reporting that 84.4% of viewers mistook an avatar for a human would replace the measured result with a different claim. The 270 counts answers from repeated ratings, not unique viewers.

A second task asked whether one clip shown alone looked real. With the person’s own voice, 65.1% of 304 answers labelled Avatar 6 real. With HeyGen’s cloned voice, that fell to 56.5% of 308 answers. Real footage itself was judged real in 78.9% of 612 answers. These are three distinct response pools. The score for video driven by original audio should not be carried over to a localized spot with a synthetic voice.

The report’s product comparisons ask another question again. In the shared-input photo test, raters scored clips separately on a 0–5 scale. Across comparisons with different scores, Avatar 6 won 69.6% against VEED Fabric 1.0 and 80.8% against OmniHuman v2. In a side-by-side footage comparison, 72.4% of expressed preferences favored Avatar 6 over Seedance 2.5, and 54.9% favored it over Kling 4.0. The Synthesia Personal video avatar comparison yielded 93.6% among expressed preferences. These figures belong to the specific product and test. HeyGen reports them for Avatar 6. They do not establish a realism score for Turbo or predict whether a particular client will approve a particular take.

The developer guide distinguishes the reference asset from the right to depict a person. All workspaces can use a photo of a person with that person’s permission. This route does not require a separate HeyGen consent agreement or consume a digital-twin slot. A video reference of an illustrated or animated character also works in any workspace. A realistic video reference of a person requires enterprise video-reference access, even if the person already approved the script. The reference clip must run for at least two seconds.

For that enterprise reference-video route, the person ordinarily verifies consent with HeyGen before submission. HeyGen says some contracts can waive its platform verification step on a case-by-case basis. Digital twins have a separate consent process. The ordinary flow asks the subject for an on-camera statement with a session-specific code and checks the subject against the training footage. The developer guide Avatar Consent also documents enterprise arrangements for submitting a recorded consent clip and separately contracted waivers. A waived platform step does not grant a producer someone else’s likeness rights.

For a customer testimonial, ask the speaker to approve the final words, the languages, the voice source and the intended placement. Permission for an English training clip may have a narrower scope than a new paid campaign that makes the same person appear to endorse a product in another language. Keep that approval with the asset or look ID and the approved audio version. A stock look can be a better fit when no individual is being represented as the actual customer.

07 — Production limitsHow long can the finished clip be?

The research report describes a model with no fixed generation-length limit and illustrates four continuous outputs lasting 25 to 28 minutes. The Model API documentation says the supplied speech determines the duration. Its two-second minimum applies to a reference video, not to the narration or the rendered clip. The model guide does not give a universal maximum duration per API request. For an hour-long course, confirm the limit and processing time for the intended workspace before promising a one-take delivery.

Product plans have their own limits. HeyGen’s November 19 public plan comparison lists an Avatar 6 maximum of one minute per video on Free and 30 minutes on Creator and Pro. The web help guide lists Turbo for Creator, both modes on Pro and Business, and a Pro trial within the Free video allowance. Those are web-plan conditions. They neither set the pay-as-you-go API price nor establish what an enterprise contract allows.

A long presenter sequence is also an editing decision. If a compliance officer changes one line in minute 18, a 25-minute source clip can be expensive to replace. Split a course at natural topic changes, keep the approved narration for each section and add captions and screen recordings in your editor. A job marked completed gives a signed video URL. Download the approved render into the production library and retain the returned video ID, width, height and duration. The status endpoint can issue a fresh URL if the signed one expires. Assembling video from reusable sections

A useful acceptance test

Ask the presenter to review the first and last ten seconds, a mid-sentence pause and the first line after a cut. Then watch the clip once without audio at its delivery size. That makes it easier to catch a gesture covering a subtitle or an expression that changes the meaning of an otherwise accurate translation.

Production choice

Approve the presenter, the frame and one take

Use a permitted portrait and finished narration when the team wants a simple generated presenter. Choose an eligible avatar look when the same person has to carry a series. A real shoot remains the right brief for a physical demonstration or an authentic unscripted exchange. For the generated route, price a small comparison and the likely retakes, then approve the frame and voice before sending the rest of the series. Our content production services cover planning and adapting a series across channels.

Digital Applied

Plan a presenter series across markets

Build repeatable approvals for scripts, voice tracks, video formats and finished edits.

LocalizationVideo formatsEditorial review
Production inputs

Start with one approved script

  • →Identify the presenter and permitted uses
  • →Choose the target frame for each channel
  • →Compare two takes at the final display size
Answers

Frequently asked questions

Both accept the same reference and speech options on the Model API, with 1080p or 2K output. Turbo costs less per second. HeyGen recommends the full model for its highest-quality performance. The vendor’s published realism study does not give Turbo the full model’s score.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source