Key Takeaways
- • Artificial Analysis ranked Claude Opus 5.5 first among 27 AI models on its new Code2Video-AA benchmark, with an Elo rating of 1536 and the top score on all five scoring axes.
- • Claude Sonnet 5.5 placed second at 1443, GPT-6 Astra third at 1427, and GPT-6.1 Sol fourth at 1369 at roughly $0.31 per task.
- • Leading open-weight models scored about 1052 to 1053 Elo, roughly 480 points behind Opus.
- • Per-task costs ranged from $0.01 to $3.89 across the evaluation, so the best choice for a team depends on volume, budget and required polish.
- • The benchmark covers 114 product and brand marketing briefs in one domain and relies on blind human preference, which the publisher notes captures taste as well as technical correctness.
Summary
Artificial Analysis launched Code2Video-AA on October 26, 2026, a benchmark that asks AI models to turn design briefs into finished, self-contained marketing videos. Models do not paint pixels directly; they write HyperFrames HTML compositions and animate them with GSAP, a JavaScript animation library. The first leaderboard covers 27 models across 114 briefs, and three tasks from the public split of the underlying Code2Video Bench are published with every model's rendered video for side-by-side comparison, according to Crypto Briefing.
Scoring is based on blind pairwise comparisons, in which evaluators view two videos without knowing which model produced them and select the better one. Head-to-head results are converted into Elo ratings across five axes: Engagement, Prompt Intent, Composition, Temporal Consistency and Craft. The five axis scores are averaged into the headline Elo figure.
Claude Opus 5.5 led the leaderboard with 1536 and topped all five axes individually. Claude Sonnet 5.5 took second at 1443, GPT-6 Astra third at 1427 and GPT-6.1 Sol fourth at 1369, with an approximate cost of $0.31 per task. Anthropic therefore holds the top two positions, while OpenAI's models fill third and fourth. The publisher describes the top three as a distinct tier above the rest of the field.
Open-weight models, whose parameters are publicly released so they can be downloaded and run by anyone, scored around 1052 to 1053 Elo, roughly 480 points behind Opus. Cost varied widely, with task prices ranging from $0.01 to $3.89 per output, and Artificial Analysis charts Elo against cost per task.
The benchmark differs from a framework developed by Show Lab, which uses code-driven generation for educational videos; Artificial Analysis targeted marketing content specifically. For companies evaluating AI tools for marketing production, the results offer a comparison on a concrete use case. The publisher cautions that the benchmark rests on 114 briefs in a single domain and that blind human preference reflects taste as much as technical correctness.
Scoring is based on blind pairwise comparisons, in which evaluators view two videos without knowing which model produced them and select the better one. Head-to-head results are converted into Elo ratings across five axes: Engagement, Prompt Intent, Composition, Temporal Consistency and Craft. The five axis scores are averaged into the headline Elo figure.
Claude Opus 5.5 led the leaderboard with 1536 and topped all five axes individually. Claude Sonnet 5.5 took second at 1443, GPT-6 Astra third at 1427 and GPT-6.1 Sol fourth at 1369, with an approximate cost of $0.31 per task. Anthropic therefore holds the top two positions, while OpenAI's models fill third and fourth. The publisher describes the top three as a distinct tier above the rest of the field.
Open-weight models, whose parameters are publicly released so they can be downloaded and run by anyone, scored around 1052 to 1053 Elo, roughly 480 points behind Opus. Cost varied widely, with task prices ranging from $0.01 to $3.89 per output, and Artificial Analysis charts Elo against cost per task.
The benchmark differs from a framework developed by Show Lab, which uses code-driven generation for educational videos; Artificial Analysis targeted marketing content specifically. For companies evaluating AI tools for marketing production, the results offer a comparison on a concrete use case. The publisher cautions that the benchmark rests on 114 briefs in a single domain and that blind human preference reflects taste as much as technical correctness.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.