Claude Opus 5.5 Leads Code-to-Video Test at $3.08 Per Task

Claude Opus 5.5 led a 27-model evaluation of AI-generated marketing videos built from code, while GPT-6.1 Sol ranked fourth at about one-tenth of Opus 5.5’s estimated cost per task.

Results dated Oct. 26 put Opus at a mean Elo score of 1,536 ±26. Claude Sonnet 5.5 scored 1,443 ±24, and GPT-6 Astra scored 1,427 ±24. The test did not establish a clear margin between those two results. GPT-6.1 Sol recorded 1,369 ±23. Opus placed first in all five areas: engagement, prompt intent, composition, temporal consistency and craft.

Code2Video-AA evaluated a specific production workflow. Each of the 27 models received one brief and HeyGen’s system prompt, without tools, then returned an HTML composition. HTML defined the visual layout, while GSAP supplied the animation. HyperFrames converted the code into a video. The briefs called for short product and brand clips.

Gemini 4 Argon compared anonymized videos in pairs across the five scoring areas. The overall Elo was the mean of those five area scores, anchored to GPT-5.5 (xhigh) at 1,000. The reported uncertainty reflects sampling of video comparisons; it does not measure variation among briefs or repeated runs of a model.

The estimated average cost per task was $3.08 for Opus 5.5 and $0.31 for GPT-6.1 Sol. Across the evaluated models, estimated task costs ranged from $0.01 to $3.89. These figures describe the cost of completing a task in this evaluation.

The test used 114 of the 168 briefs in HeyGen’s original benchmark: 39 from the public split and 75 private briefs shared for the evaluation. That sample and the code-based workflow define the scope of the results.

This benchmark measures marketing-video creation, a different task from multi-step agent work. An earlier agent-ranking evaluation of Opus 5.5 assessed tasks through outcomes and feedback from real user sessions.