Artificial Analysis launches Code2Video-AA

Artificial Analysis published Code2Video-AA results for 27 model configurations across 114 video-design briefs. Its evaluation compares rendered code compositions using model judging, rather than measuring direct video generation.

Videos made by writing code

Artificial Analysis launched its independent implementation of HeyGen's Code2Video Bench on October 26. A model receives a design brief and returns an HTML composition using the HyperFrames format; rendered videos are compared by a Gemini 4 Argon judge. The launch covers 27 model configurations and 114 briefs, including 39 public briefs and 75 private briefs shared by HeyGen. [1][3]

The resulting Elo rankings compare this particular code-to-video workflow, rather than direct video generation. Each of five judging axes has its own comparisons and Elo fit, and the overall score averages those axes. Render failures lose against successfully rendered compositions. [1]

Sources

  1. Code2Video-AA: how well do AI models design videos in code?

    Artificial Analysis · Oct 26, 2026

  2. Code2Video-AA benchmarks 27 models on code to video

    bridgecode · Oct 27, 2026

  3. Artificial Analysis Intelligence Benchmarking Methodology

    Artificial Analysis · retrieved Nov 3, 2026