Claude Opus 5.5 tops code-to-video benchmark
On Artificial Analysis' new Code2Video-AA leaderboard, Claude Opus 5.5 leads by 93 Elo at designing short brand videos in code, though cheaper models offer much of that quality for far less money.
Artificial Analysis has launched Code2Video-AA, its own independent implementation of HeyGen's Code2Video Bench. The benchmark tests how well models write short product and brand videos as code, working from 114 design briefs.
Each model gets a single call containing the brief and HeyGen's system prompt, with no tools. It must return one HyperFrames HTML composition that uses HTML for layout and GSAP for animation. The rendered videos are compared in pairs by a judge on five axes, each with its own Elo rating. The overall score is the average of the five.
Results. Claude Opus 5.5 (max with fallback) leads at 1536 Elo, 93 points ahead of the next model, and ranks first on all five axes. Its biggest margin is on craft (128 points) and its smallest is on prompt intent (48). Claude Sonnet 5.5 (1443) and GPT-6 Astra (1427) are statistically tied for second. GPT-6.1 Sol (1369) is fourth.
Cost. Cost per task ranges from $0.01 (MiMo-V2.6-Flash) to $3.89 (Claude Fable 5.1), a spread of more than 300x. GPT-6.1 Sol costs $0.31 per task, about a tenth of Opus 5.5's $3.08. GPT-6 Astra costs $1.62, five times as much as Sol, for 58 more Elo points. Most of the gap in output tokens comes from reasoning: the three Claude 5.5 models use about 152k–176k tokens per task, while the median for the other models is about 34k.
Open weights. The best open-weights models, DeepSeek V4.1 Flash, MiMo-V2.6-Flash and MiMo-V2.6-Pro, score around 1052–1053, at least 483 points behind the leader.
Render reliability was high: 98.8% of the 3,078 compositions rendered, and every one of the 27 models rendered at least 96% of its videos.