artificial-analysis-leaderboard

Artificial Analysis Leaderboard (Live) — Agent Skill

A portable coding-agent skill that fetches live LLM model data from artificialanalysis.ai/leaderboards/models — including Artificial Analysis' three composite indices:

• Intelligence Index (v4.1)
• Coding Index (SWE-Bench Verified / LiveCodeBench)
• Agentic Index (τ²-Bench / GDPval-AA / Terminal-Bench)

…plus median output speed, time-to-first-token, and blended pricing. ~225 models, continuously updated.

The 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.

The 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).

What it returns

-- 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).

Usage

# Top 15 models by Intelligence Index (default sort)
node aa-leaderboard.js --top 15

# Top 10 by CODING
node aa-leaderboard.js --sort coding --top 10

# Top 10 by AGENTIC (tool use / agents)
node aa-leaderboard.js --sort agentic --top 10

# Top 10 by SPEED
node aa-leaderboard.js --sort speed --top 10

# Cheapest models (price ascending)
node aa-leaderboard.js --sort price --top 10

# Lowest latency (snappiest first-chunk)
node aa-leaderboard.js --sort latency --top 10

# Filter by creator
node aa-leaderboard.js --creator anthropic --top 10
node aa-leaderboard.js --creator deepseek

# Filter by model name (substring, case-insensitive)
node aa-leaderboard.js --model 'mini' --top 10

# Deep profile of ONE model: identity, all 3 indices, every benchmark,
# full price breakdown (input/output/cache/blends), and the speed/latency
# percentile distribution (p5–p95). Works by display name OR slug.
node aa-leaderboard.js --deep --model 'Claude Opus 5 (max)'
node aa-leaderboard.js --deep --model claude-opus-5          # slug works too
node aa-leaderboard.js --deep --model claude-opus-5 --json   # raw ~90-field payload

# Machine-readable JSON (for piping / further processing)
node aa-leaderboard.js --json --top 20
node aa-leaderboard.js --json --sort coding --creator google

Flags

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.

Deep single-model profile (--deep)

--deep trades the many-rows table for a full profile of one model, surfacing the ~90 fields the site embeds per model (the table shows only ~8). Requires --model X (name or slug, substring OK; an exact match wins ties). Sections:

• Identity — full name, slug, creator + country, release date, context window, size/price tier, reasoning / open-weights flags, modalities, OpenRouter id
• Composite Indices — Intelligence / Coding / Agentic (0–100)
• 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)
• Pricing — input / output / cache-read (w/ discount %) / cache-write, plus blended variants (7:2:1 is the leaderboard's headline price)
• 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
• Intelligence Index internals — cost / time / tokens per eval task

--deep --json emits the whole decoded object (nested objects intact, $undefined normalised to null) for programmatic use.

Install as an agent skill

Agents discover this skill by finding SKILL.md inside <skills-dir>/artificial-analysis-leaderboard/. Clone into whichever skills directory your agent scans:

# Example locations (use the one for your agent):
#   gjc : ~/.gjc/agent/skills/
#   pi  : ~/.pi/agent/skills/
git clone https://github.com/pinion05/artificial-analysis-leaderboard.git \
  <your-skills-dir>/artificial-analysis-leaderboard

Restart your agent session, then trigger it with /skill:artificial-analysis-leaderboard, or ask in natural language:

• "요즘 가장 똑똑한 모델 뭐야?" / "best LLM right now"
• "코딩 잘하는 모델", "에이전트 점수", "model leaderboard"
• "가장 빠른 모델", "가장 싼 모델"

Standalone (no agent)

The script is fully self-contained — it works anywhere with Node.js 14+:

curl -O https://raw.githubusercontent.com/pinion05/artificial-analysis-leaderboard/main/aa-leaderboard.js
node aa-leaderboard.js --sort coding --top 10

How it works

The 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.

Files

• aa-leaderboard.js — fetch + parse script (Node.js stdlib, no deps)
• SKILL.md — portable agent skill definition
• README.md — this file

Requirements

• Node.js 14+ (uses only https from the standard library)
• Internet access to artificialanalysis.ai

Reliability

• Plain HTTPS GET with a browser User-Agent. If the site changes its embedded payload shape, the parser's parse() / field() will need updating.
• Data is live at fetch time; Artificial Analysis updates continuously.
• 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.)

License

MIT