Photo: Tim Witzdam / Pexels
Artificial Analysis tested 27 models on 114 design briefs, and the gap between proprietary leaders and open-weight models was wide
Artificial Analysis has a new way to grade AI models: ask them to make a marketing video, then let people judge the results blind.
The benchmark is called Code2Video-AA, and its first leaderboard covers 27 models. Claude Opus 5.5 finished on top with an Elo rating of 1536, ahead of every rival on every scoring axis.
How the test works #
Artificial Analysis launched Code2Video-AA on October 26, 2026. The setup asks each model to turn a design brief into a finished, self-contained video.
The catch is that no model paints pixels directly. Each one writes code instead, building HyperFrames HTML compositions and animating them with GSAP, a JavaScript animation library.
The briefs number 114, all centered on product and brand marketing. Three tasks from the public split of the underlying Code2Video Bench are posted with every model’s rendered video, so anyone can compare outputs side by side.
Scoring relies on blind pairwise comparisons. Evaluators see two videos without knowing which model made them, then pick the better one.
Those head-to-head results are converted into Elo ratings for each of five axes:
- Engagement : does the video hold attention
- Prompt Intent : does it deliver what the brief asked for
- Composition : is the layout visually coherent
- Temporal Consistency : does it stay stable from frame to frame
- Craft : how polished the execution is
The five per-axis scores are then averaged into the headline Elo number.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The leaderboard #
Claude Opus 5.5 led with 1536. Claude Sonnet 5.5 took second at 1443, and GPT-6 Astra came third at 1427.
GPT-6.1 Sol landed in fourth with an Elo of 1369. Its cost came in at approximately $0.31 per task.
Opus also topped all five axes individually.
Cost varied widely across the evaluation, with task prices ranging from $0.01 to $3.89 per output. Artificial Analysis also charts Elo against cost per task.
Open-weight models trail far behind #
The leading open-weight models scored around 1052 to 1053 Elo. That puts them roughly 480 points behind Opus.
Open-weight models are those whose underlying parameters are publicly released, so anyone can download and run them.
Why this benchmark is different #
It also differs from a framework developed by Show Lab, which uses code-driven generation for educational videos. Artificial Analysis aimed its version squarely at marketing content.
What this means #
For companies weighing AI tools for marketing production, the leaderboard offers a comparison on a real use case. Opus is the clear quality leader, and the top three form a distinct tier above everyone else. With per-task prices running from $0.01 to $3.89, the best model for a given team depends on volume, budget, and how much polish a campaign needs. GPT-6.1 Sol at approximately $0.31 per task shows that a top-four result does not have to come at a premium.
Anthropic holds the top two spots, while OpenAI’s models fill out third and fourth. There are limits worth keeping in mind. The benchmark rests on 114 briefs in one domain, and blind human preference captures taste as much as technical correctness.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our