Code Arena has evolved rapidly alongside model capabilities. Early interactions often involved single-file HTML pages or small frontend components. Today, users ask models to build multi-file React applications, implement dashboards, develop consumer products, create browser games, simulate dynamic systems, and build editing tools.
The evaluation setting has also changed. Models increasingly operate with tools for file editing, execution, web access, and media understanding. As a result, Code Arena is no longer only measuring whether models can write valid code. It is measuring whether models can complete diverse, product-oriented development tasks under realistic conditions.
This shift makes a single global leaderboard less sufficient. Web development tasks vary substantially in intent and required capability: a polished landing page, a data-heavy admin dashboard, a browser game, and a physics simulation stress different model behaviors. Category-specific leaderboards provide a more interpretable way to measure these differences.
Building the Taxonomy
We analyzed more than 250,000 filtered Code Arena prompts collected over five months, focusing on web development tasks. We used clustering analysis to identify recurring patterns in user intent and task structure, then refined the resulting groups through an iterative taxonomy-building process. The refinement process optimized for four criteria:
- Interpretability: Categories should be understandable to users, researchers, and model developers. Each category should describe a recognizable type of web development task or a representative user intent.
- Coverage: The taxonomy should capture a broad share of real user behavior in Code Arena.
- Statistical robustness: Each category should contain enough examples to support reliable leaderboard estimates.
- Boundary clarity: Categories should distinguish meaningful differences in task intent and model capability, while still allowing overlap when prompts are naturally multi-domain.
