LMArena AI Model Leaderboard

Problem

When comparing AI model performance across text, code, vision, search, documents, and multimodal generation, raw leaderboard files are split by category and hard to inspect in one place.

How It Works

The skill fetches data from oolong-tea-2026/arena-ai-leaderboards into arena-data.json. If the script fails, it can fall back to curl requests for latest.json or dated/category files. It then renders a static HTML page with horizontal tabs for 10 subleaderboards, showing Elo scores, API pricing, LobeHub icons, and a responsive table. China-based models are marked with ch: true using vendor rules such as Alibaba, ByteDance, Z.ai, Moonshot, Baidu, MiniMax, DeepSeek, and Xiaomi, highlighted in red with a 🇨🇳 tag; vendors like Anthropic, OpenAI, Google, and xAI should not be mislabeled. Pricing uses built-in $/million-token values, while newer models can be checked on llm-stats.com. If dated files return 404, use latest.json; if the Node script fails, inspect generate-html.js and rebuild the page structure manually.

Boundaries

This is best for generating a local, read-only ranking page rather than full model evaluation or a production service. Coverage depends on the upstream repository providing files such as text.json, code.json, and vision.json; rankings and model availability reflect source data.