Code2Video-AA evaluates AI models on video creation from code

HeyGen: HeyGen develops AI tools focused on video creation and marketing content generation. It originated the Code2Video Bench that evaluates models on turning product and brand briefs into video compositions. Artificial Analysis built Code2Video-AA in partnership with HeyGen to provide an independent implementation of this benchmark. Claude Opus: Claude Opus 5.5 is Anthropic’s leading frontier model, optimized for high performance in complex reasoning and creative generation tasks. It tops the Code2Video-AA leaderboard with the highest overall Elo score and leads on every evaluation axis including engagement, composition, and craft. The model’s outputs are rendered via HyperFrames HTML for video assessment. GPT-6 Astra: GPT-6 Astra is an OpenAI model variant emphasizing advanced performance in complex tasks. It ties for second on the Code2Video-AA leaderboard alongside Claude Sonnet 5.5. Its results show competitive video composition abilities relative to higher-cost leaders. GPT-6.1 Sol: GPT-6.1 Sol is an OpenAI model variant designed with emphasis on cost efficiency alongside strong capabilities. It ranks fourth on Code2Video-AA with a competitive Elo while incurring far lower per-task costs than leading models. Its performance demonstrates effective trade-offs in video code generation tasks. HyperFrames: HyperFrames is a system for building animated HTML compositions using layout code and GSAP animations that can be rendered into videos. Models in Code2Video-AA generate HyperFrames index.html files from design briefs, which are then converted to video for judging with no external network access. Code2Video-AA: Code2Video-AA is Artificial Analysis’ independent benchmark testing AI models on creating short product and brand marketing videos from code. It evaluates 114 design briefs by having models output HyperFrames HTML compositions that are rendered and ranked via Elo scores from Gemini 4 Argon judgments on engagement, prompt intent, composition, temporal consistency, and craft. Gemini 4 Argon: Gemini 4 Argon is a Google model used in the Code2Video-AA evaluation both as a judge and as a participating model on the leaderboard. It performs pairwise comparisons of rendered videos across five specific axes to determine relative quality. The model itself ranks sixth overall in the benchmark. MiMo-V2.6-Flash: MiMo-V2.6-Flash is an open-weights model optimized for cost-effective inference in multimodal and generative tasks. It ties for the highest score among open models on Code2Video-AA while maintaining very low per-task costs. Its results highlight the gap between open and closed models on creative video benchmarks. Claude Sonnet 5.5: Claude Sonnet 5.5 is Anthropic’s balanced mid-tier model offering strong capabilities across reasoning and generation. It ties statistically for second place on Code2Video-AA overall and performs consistently across the five judging axes. The model uses fallback configurations in the benchmark runs. Artificial Analysis: Artificial Analysis is an independent organization specializing in AI model evaluations and benchmarks across capabilities like reasoning, coding, and agentic tasks. It developed Code2Video-AA as its own implementation of HeyGen’s Code2Video Bench to assess how well models generate marketing videos in code. The firm partners with external organizations to create transparent, standardized testing frameworks using Elo ratings from pairwise judgments. DeepSeek V4.1 Flash: DeepSeek V4.1 Flash is a high-efficiency open-weights model from DeepSeek focused on strong performance at low cost. It leads the open-weights group on Code2Video-AA but trails the top proprietary models by a wide margin. The model participates in single-response evaluations without tools. Benchmarking: Independent implementations of specialized benchmarks enable transparent comparisons of AI models on creative video generation tasks using code outputs. Cost Efficiency: AI models exhibit wide variation in the trade-off between performance and per-task cost when evaluated on video design benchmarks. Open Weights Performance: Open-weights models demonstrate consistent but substantial gaps behind leading proprietary models in evaluations of code-based marketing video creation.