Artificial Analysis has introduced the Code2Video-AA benchmark, which evaluates how AI models generate marketing videos based on design briefs. The testing process involves 27 models and 114 specific tasks, where each AI writes code using HyperFrames and GSAP to render final video content.
The evaluation relies on blind pairwise comparisons, with results converted into Elo ratings. Claude Opus 5.5 secured the top position with an Elo rating of 1536, followed by Claude Sonnet 5.5 at 1443 and GPT-6 Astra at 1427. GPT-6.1 Sol ranked fourth with an Elo of 1369.
The cost per task varies significantly, ranging from $0.01 to $3.89. For instance, the fourth-ranked GPT-6.1 Sol costs approximately $0.31 per task. Leading open-weight models scored between 1052 and 1053 Elo, trailing the top proprietary models by roughly 480 points.
The benchmark focuses exclusively on product and brand marketing content. While Anthropic currently occupies the top two spots and OpenAI holds the third and fourth positions, the researchers noted that human preference in these tests reflects subjective taste alongside technical performance.
