Code2Video-AA benchmarks 27 models on code to video

The benchmark analyzed 27 models across 114 design briefs. Instead of generating a video file directly, each model writes an HTML composition that is passed through a renderer before evaluation.

In this article you will see:

• How Code2Video-AA evaluates videos made with code
• Who leads the current leaderboard
• What the render rate reveals
• How to interpret the results without overgeneralizing

How Code2Video-AA turns briefs into videos

Code2Video-AA is Artificial Analysis's implementation of the public Code2Video Bench dataset, created by HeyGen. Each brief describes a short product or brand piece, and the model produces an HTML composition with HyperFrames.

The frame above shows an example output from the dataset. The workflow described by the team sends a brief and the benchmark instructions to the model in a single call, with no tools; then the code is rendered to video for evaluation.

This distinction changes what is being measured. The test includes the ability to follow visual instructions, write the composition, and produce output that the renderer can execute. It does not measure only the quality of a finished video received as input.

The current Code2Video-AA leaderboard

On the leaderboard published on October 26, Claude Opus 5.5 leads with 1536 Elo points and comes ahead on all five evaluated axes. Second place is Claude Sonnet 5.5, with GPT-6 Astra right behind.

The gap between first and second place is 93 points. Artificial Analysis reports that Opus 5.5 also leads on each axis, with its biggest advantage in craft and its smallest in prompt intent, both compared to Sonnet 5.5.

Elo is a relative measure within this benchmark, not an accuracy percentage nor a universal quality score. The page anchors GPT-5.5 in xhigh mode at 1000 points and defines that higher values indicate a higher position in this comparison.

To check the full list and the current status, see the official Code2Video-AA leaderboard. Versions, configurations, and dataset date matter when comparing numbers published at different times.

What the render rate reveals

Of the 3,078 possible results, 3,042 were rendered, a rate of 98.8%. Eight of the 27 models managed to render all 114 tasks.

GLM-5.3 in max mode had the lowest rate, with 110 renders out of 114 tasks. The three models at the bottom rendered between 111 and 113 videos, and the team attributes most of the Elo difference to judgments of the outputs that were able to render.

A rendering failure prevents the judge from evaluating the video for that task. For this reason, technical success rate and visual evaluation need to be read separately: code that does not run loses the comparison, even if the idea described in the brief seems correct.

What Code2Video-AA does not prove

The public dataset contains 114 briefs and covers short product and brand videos. It is useful for comparing models on this bounded task, but it does not represent every type of edit, duration, visual identity or audiovisual production workflow.

The run is an independent evaluation by Artificial Analysis on dedicated hardware, but it remains a single implementation with a specific set of instructions and renderer. The result does not guarantee the same ranking with another tool, another brief or a different human review.

The briefs come from the public part of HeyGen's Code2Video Bench, licensed under Creative Commons Attribution 4.0. The Artificial Analysis page also leaves examples of videos generated by the models, which lets you inspect concrete outputs beyond the aggregate score.

How to test Code2Video-AA today

Open the Code2Video-AA leaderboard and pick a few briefs that relate to the format you want to produce. Compare the videos for the same brief and look at the rendered result, not just the model's position.

If you are evaluating models in your own workflow, keep the prompt, output size and renderer the same. Record which model and configuration you used, when you ran the test and whether the composition rendered. That way a technical failure does not get mixed up with a low visual rating.

Use criteria that match your goal. A piece to promote a product may require brand adherence; a short tutorial may require clear instructions; an animation with multiple elements may depend on temporal consistency. The ranking helps choose candidates, but testing with your brief decides whether the result works.

This care applies to other rankings too. In CursorBench 4.0 with DeepSeek V4.1 Flash, the context is code editing inside Cursor. The Terminal-Bench 4.0 with Kimi K3.1 Pro measures agent tasks in terminal. Each scoreboard answers a different question.