Genativity

news

HeyGen ships Decision 1, a model that scores video from sampled frames and publishes no accuracy numbers

The new Decisions endpoint takes up to four images and one video per request and returns probabilities for named yes/no, choice and score questions, with pricing listed on OpenRouter and no benchmark results anywhere.

By Priya Shenoy ยท

HeyGen Decision 1 is a judging model from the video company HeyGen. It reads text or JSON plus up to four images and one video, then answers named yes/no, choice and score questions with a probability attached to each. The video is read as sampled frames. HeyGen publishes no accuracy figures for it.

HeyGen has released a model that does not generate anything. Decision 1 reads a state (text or JSON), optional media (up to four images and one video), and a set of named questions, then returns typed answers with a probability attached to each one. It is pitched at exactly the job this site cares about, scoring media against criteria, and the documentation that describes it carries no accuracy measurement of any kind.

Three question types, one synchronous call

The endpoint is POST /v3/models/decisions, and it exposes two models, heygen-decision-1 and heygen-decision-1-turbo, the second of which HeyGen lists as a preview. Each named question is one of three types: noul, a probability that a statement is true; choice, a pick from named options; or score, a rating on an ordered scale the caller defines. Criteria are optional for yes/no questions and required for the other two, which means the rubric for any visual quality score comes from whoever writes the request, not from HeyGen. HeyGen's model index describes the whole thing as one synchronous call with a probability for every answer, so there is no polling or webhook step.

Video is judged by frames, and the sound is the caller's problem

The video documentation is the part that matters for anyone thinking about using this to grade generated clips. A request takes one video alongside up to four images, and HeyGen states plainly that the model reads it as sampled frames, placed after any images and before the state. If you want the model to judge what is said, the instruction is to paste the transcript into the state yourself. A judge that scores a video without hearing it is a reasonable engineering tradeoff, but it is a hard limit on what a lipsync or audio sync verdict from this model could ever mean.

The worked examples in HeyGen's own docs are product and brand checks rather than model evaluation: comparing a package shown in a demo video against an approved design, scoring how clearly a logo reads, or checking a submitted thumbnail against a template and rating title legibility at small sizes. That is a short step away from scoring generated output, which is the obvious use once the clip being checked came from a model in the first place.

In our benchmark, this is the kind of generated product footage such a scorer would be pointed at.

Seedance 2.5 on our "Watch, wide then macro" test, scored 5.00/5. Same prompt, every model

A price exists, a score does not

OpenRouter's listing dates both models to November 13, 2026 and gives Decision 1 a 131,072 token context with input billed at $0.14 per million tokens and output at $0, with the Turbo variant at $0.09 per million input tokens. Sampled video frames count toward input tokens, so the cost of judging a clip scales with how much of it the model looks at. Media limits are documented too: 50 MiB per URL download and 16 MiB for a decoded base64 video inside the request body.

What is missing is any number describing how often the answers are right. There is no accuracy table, no agreement rate against human raters, and no calibration data showing whether a 0.96 probability means 96 percent of those answers hold up. For a product whose entire output is a confidence number, that absence is the story. Work like DecisionPivot's run across 40,033 items, where 19 of 40 automated judges landed under a guessing baseline, is the reason a confident-looking probability is not evidence on its own.

HeyGen has form for shipping ahead of the scoreboard, including HeyGen Video 1's three Elo lead on a silent text-to-video board that two Artificial Analysis pages scored differently. Decision 1 at least ships with a documented interface and explicit media limits. Until someone measures it, a scoring model with no scores is a tool, not a verdict.

Sources: HeyGen Documentation, HeyGen Decision from video, HeyGen Decision from images, OpenRouter.

Cite this

Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.

Genativity. (2026). HeyGen ships Decision 1, a model that scores video from sampled frames and publishes no accuracy numbers. Retrieved November 16, 2026, from https://genativity.com/news/heygen-decision-1-video-sampled-frames-no-accuracy-numbers
<a href="https://genativity.com/news/heygen-decision-1-video-sampled-frames-no-accuracy-numbers">HeyGen ships Decision 1, a model that scores video from sampled frames and publishes no accuracy numbers</a> (Genativity)

Priya Shenoy

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.