{"id":"69a99e1d-e545-4a45-9e8a-5c8805f4e96a","slug":"artificial-analysis-leaderboard","kind":"skill","name":"artificial-analysis-leaderboard","summary":"Portable agent skill: live LLM leaderboard data from artificialanalysis.ai — Intelligence, Coding & Agentic indices, speed, pricing. No API key, zero deps (Node.js stdlib).","owner":"pinion05","repoUrl":"https://github.com/pinion05/artificial-analysis-leaderboard","license":"MIT","licenseTier":"permissive","trust":"community","stars":4,"match":null,"fewStars":true,"format":"skill","createdAt":"2026-07-09T13:17:18.000Z","addedAt":"2026-10-21T23:26:08.252Z","logoUrl":"https://avatars.githubusercontent.com/u/117980824?u=7c1e1b0b178683772e7c2f18454944b538c52da3&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["data-analytics"],"tags":[],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-26T20:02:42.189Z","sample":false,"url":"https://conduits.dev/tools/artificial-analysis-leaderboard","language":"JavaScript","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"# Artificial Analysis Leaderboard (Live) — Agent Skill\n\nA portable **coding-agent skill** that fetches **live** LLM model data from [artificialanalysis.ai/leaderboards/models](https://artificialanalysis.ai/leaderboards/models) — including Artificial Analysis' three composite indices:\n\n- **Intelligence Index** (v4.1)\n- **Coding Index** (SWE-Bench Verified / LiveCodeBench)\n- **Agentic Index** (τ²-Bench / GDPval-AA / Terminal-Bench)\n\n…plus median output speed, time-to-first-token, and blended pricing. ~225 models, continuously updated.\n\nThe skill follows the common `SKILL.md` format (a `name` + `description` frontmatter and a Markdown body), so it drops into **any agent that loads skills from a directory** — gjc, pi, and similar. The fetch script is fully self-contained.\n\nThe leaderboard page is server-side rendered (Next.js) and embeds the full per-model data payload in the HTML, so the script extracts that directly — **no API key, no JS rendering, no scraping of dynamic content, zero dependencies** (Node.js stdlib only).\n\n## What it returns\n\n| Field | Meaning |\n|---|---|\n| `model` | Model name (e.g. `GPT-5.5 (xhigh)`) |\n| `creator` | Vendor (Anthropic, OpenAI, Google, DeepSeek, …) |\n| `intelligence` | **Intelligence Index v4.1** — main quality score (higher = better). Trailing `*` = estimated. |\n| `coding` | **Coding Index** (higher = better at code) |\n| `agentic` | **Agentic Index** (higher = better at tool use / agents) |\n| `blendedUSD` | Blended price per 1M tokens (lower = cheaper) |\n| `speedTokPerSec` | Median output throughput, tokens/sec (higher = faster) |\n| `latencyFirstChunk` | Time to first chunk, seconds (lower = snappier) |\n| `totalResponse` | End-to-end response time, seconds |\n\n`--` means that model has no data for that metric (e.g. a newly released model, or one not yet benchmarked on the coding/agentic suites).\n\n## Usage\n\n```bash\n# Top 15 models by Intelligence Index (default sort)\nnode aa-leaderboard.js --top 15\n\n# Top 10 by CODING\nnode aa-leaderboard.js --sort coding --top 10\n\n# Top 10 by AGENTIC (tool use / agents)\nnode aa-leaderboard.js --sort agentic --top 10\n\n# Top 10 by SPEED\nnode aa-leaderboard.js --sort speed --top 10\n\n# Cheapest models (price ascending)\nnode aa-leaderboard.js --sort price --top 10\n\n# Lowest latency (snappiest first-chunk)\nnode aa-leaderboard.js --sort latency --top 10\n\n# Filter by creator\nnode aa-leaderboard.js --creator anthropic --top 10\nnode aa-leaderboard.js --creator deepseek\n\n# Filter by model name (substring, case-insensitive)\nnode aa-leaderboard.js --model 'mini' --top 10\n\n# Deep profile of ONE model: identity, all 3 indices, every benchmark,\n# full price breakdown (input/output/cache/blends), and the speed/latency\n# percentile distribution (p5–p95). Works by display name OR slug.\nnode aa-leaderboard.js --deep --model 'Claude Opus 5 (max)'\nnode aa-leaderboard.js --deep --model claude-opus-5          # slug works too\nnode aa-leaderboard.js --deep --model claude-opus-5 --json   # raw ~90-field payload\n\n# Machine-readable JSON (for piping / further processing)\nnode aa-leaderboard.js --json --top 20\nnode aa-leaderboard.js --json --sort coding --creator google\n```\n\n### Flags\n\n| Flag | Description |\n|---|---|\n| `--top N` | Limit to top N rows (after sort + filter) |\n| `--sort FIELD` | `intelligence` (default) · `coding` · `agentic` · `speed` · `price` · `latency` · `name` |\n| `--creator X` | Filter: creator contains X (case-insensitive) |\n| `--model X` | Filter: model name **or slug** contains X (case-insensitive). Required with `--deep`. |\n| `--deep` | Full single-model profile (use with `--model X`): indices, every benchmark, price breakdown, latency percentiles |\n| `--table` | Human-readable aligned table (default) |\n| `--json` | Emit JSON array; with `--deep`, emit the model's full ~90-field object |\n\n**Sort direction:** `intelligence`/`coding`/`agentic`/`speed` sort **descending** (best first); `price`/`latency` sort **ascending** (cheapest/snappiest first); `name` sorts A→Z. Rows with `--` in the sort field sink to the bottom.\n\n## Deep single-model profile (`--deep`)\n\n`--deep` trades the many-rows table for a **full profile of one model**, surfacing\nthe ~90 fields the site embeds per model (the table shows only ~8). Requires\n`--model X` (name or slug, substring OK; an exact match wins ties). Sections:\n\n- **Identity** — full name, slug, creator + country, release date, context window, size/price tier, reasoning / open-weights flags, modalities, OpenRouter id\n- **Composite Indices** — Intelligence / Coding / Agentic (0–100)\n- **Benchmarks** — GPQA, MMMU-Pro, HLE, Terminal-Bench (v2.1 + hard), SciCode, τ²-Bench, Long-Context Reasoning, GDPval (normalized + Elo w/ 95% CI), Critical Point, Analyst Agent, IT-Bench (SRE), IF-Bench, Omniscience (accuracy / non-hallucination / score)\n- **Pricing** — input / output / cache-read (w/ discount %) / cache-write, plus blended variants (7:2:1 is the leaderboard's headline price)\n- **Speed & latency distribution** — throughput and time-to-first-token across **p5 / p25 / p50 / p75 / p95** (a single median hides the long-tail reasoning runs), plus end-to-end and answer speed\n- **Intelligence Index internals** — cost / time / tokens per eval task\n\n`--deep --json` emits the whole decoded object (nested objects intact, `$undefined` normalised to null) for programmatic use.\n\n## Install as an agent skill\n\nAgents discover this skill by finding `SKILL.md` inside `<skills-dir>/artificial-analysis-leaderboard/`. Clone into whichever skills directory your agent scans:\n\n```bash\n# Example locations (use the one for your agent):\n#   gjc : ~/.gjc/agent/skills/\n#   pi  : ~/.pi/agent/skills/\ngit clone https://github.com/pinion05/artificial-analysis-leaderboard.git \\\n  <your-skills-dir>/artificial-analysis-leaderboard\n```\n\nRestart your agent session, then trigger it with `/skill:artificial-analysis-leaderboard`, or ask in natural language:\n\n- \"요즘 가장 똑똑한 모델 뭐야?\" / \"best LLM right now\"\n- \"코딩 잘하는 모델\", \"에이전트 점수\", \"model leaderboard\"\n- \"가장 빠른 모델\", \"가장 싼 모델\"\n\n### Standalone (no agent)\n\nThe script is fully self-contained — it works anywhere with Node.js 14+:\n\n```bash\ncurl -O https://raw.githubusercontent.com/pinion05/artificial-analysis-leaderboard/main/aa-leaderboard.js\nnode aa-leaderboard.js --sort coding --top 10\n```\n\n## How it works\n\nThe leaderboard page is Next.js SSR; the full per-model data payload is embedded in the HTML as escaped JSON. `parse()` splits on each model's `id`, `field()` reads each field by its escaped key, and records are de-duplicated by slug (preferring the copy with the most non-null metrics). `--deep` goes further: `deepExtract()` locates the target record, `braceMatch()` walks from its opening brace to its matching close (treating `\\\"` as an in-string toggle so braces inside string values are skipped), and the slice is double-parsed — once to unescape the embedded JSON string, once to materialise the object — yielding the full ~90-field model payload. The blended price uses Artificial Analysis' `price1mBlended7To2To1` blend, which matches the site's default table column.\n\n## Files\n\n- `aa-leaderboard.js` — fetch + parse script (Node.js stdlib, no deps)\n- `SKILL.md` — portable agent skill definition\n- `README.md` — this file\n\n## Requirements\n\n- Node.js 14+ (uses only `https` from the standard library)\n- Internet access to `artificialanalysis.ai`\n\n## Reliability\n\n- Plain HTTPS GET with a browser User-Agent. If the site changes its embedded payload shape, the parser's `parse()` / `field()` will need updating.\n- Data is live at fetch time; Artificial Analysis updates continuously.\n- No API key required — public leaderboard. (Artificial Analysis also offers a paid Data API for individual benchmark breakdowns and time series; this skill does not use it.)\n\n## License\n\nMIT\n"}