artificialanalysis-ai-parser

Parser for artificialanalysis.ai — extracts AI model data (pricing, benchmarks, speed) without an API key.

Why?

The idea started from demianarc/artificialanalysisscrapper — a Python scraper that fetched model data from the Artificial Analysis Next.js RSC endpoint. It was a clever approach: the site's React Server Components stream exposed the full dataset (hostsModels) in a single 10 MB response, no authentication needed.

However, after the site's redesign ("A new look for Artificial Analysis"), the old line-based parser broke completely. The RSC format changed from simple key:value pairs to a chunk-referenced wire format with I[...] inline references and $c:props:... circular links.

This project:

• Rewrites the extraction using regex + bracket-counting instead of line-based parsing
• Deduplicates 867 host-model pairs down to 326 unique models (keeping first occurrence with full non-circular data)
• Cleans the output to only essential fields (pricing, IQ, speed, context window)
• Outputs models.json — 314 models with full input/output/cache pricing, ready for downstream use

The result is a self-contained Python script with zero dependencies beyond the standard library.

Quick start

C++

g++ -std=c++17 -O2 artificialanalysis.ai-parser.cpp -lcurl -o aaparser
./aaparser --minimal --pretty          # fetch + save to models.json

Requires: libcurl, nlohmann/json (header-only, auto-downloaded if missing).

Python

python3 artificialanalysis.ai-parser.py --minimal --pretty

JavaScript (Node.js)

// Node.js — works without CORS restrictions
const { AAParser } = require('./artificialanalysis.ai-parser.js');
const models = await AAParser.fetch({ minimal: true });
console.log(models[0].name, models[0].price_1m_input_tokens);

Note: The JS parser does not work directly in the browser. The RSC endpoint requires the custom rsc header which triggers a CORS preflight, and the server does not return Access-Control-Allow-Headers. Use in Node.js or through a CORS proxy.

Output

Downloading RSC data from https://artificialanalysis.ai/leaderboards/providers?_rsc=hgvan ...
Downloaded 10,481,155 bytes
Extracted 867 raw entries (host-model pairs)
Deduplicated to 326 unique models
Models with pricing: 314

Saved 314 models to models.json (134,549 bytes)

Top model: GPT-5.5 (xhigh) (OpenAI)
  IQ: 60.24 | Coding: 59.12 | Math: None
  Price: $5.00 in / $30.00 out
  Speed: 57 tok/s

models.json structure

Each entry:

Field	Description
name	Model name
creator	AI lab / company
slug	URL-friendly identifier
intelligence_index	AA Intelligence Index score
coding_index	AA Coding Index score
math_index	AA Math Index score
price_1m_input_tokens	Input price per 1M tokens (USD)
price_1m_output_tokens	Output price per 1M tokens (USD)
price_1m_cache_hit	Cache hit price per 1M tokens (USD)
blended_price_3_1	Blended price at 3:1 input:output ratio
context_window_tokens	Context window size
output_tokens_per_second	Generation speed
time_to_first_token_ms	Latency to first token
reasoning	Whether it's a reasoning model
open_weights	Whether weights are open

Data coverage

Metric	Coverage
Pricing (input/output)	100% (314/314)
Intelligence Index	87%
Coding Index	90%
Math Index	60%
Speed (tok/s)	100%
Cache pricing	33%

How it works

artificialanalysis.ai
  └─ /leaderboards/providers?_rsc=hgvan
       └─ Next.js RSC stream (10 MB, text/x-component)
            └─ Contains "hostsModels":[{...}] with ~867 entries
                 └─ Extract JSON via bracket-counting
                      └─ Deduplicate by model_id
                           └─ Clean & output models.json

The RSC endpoint requires specific headers (rsc: 1, next-router-state-tree, next-url) but no cookies or authentication.

Limitations

• No API key = fragile. The RSC endpoint is an internal Next.js mechanism. If the site changes its chunk format again, the bracket-counting may need updating.
• Circular references. From the 2nd entry onward, some nested model fields use $c:props:... reference strings instead of actual values. We keep only the first occurrence per model_id (which has full data).
• Official API is preferred for production use. This parser is a workaround for when you don't have (or don't want) an API key. See artificialanalysis.ai/documentation for the free API tier (1,000 req/day).

Companion: interactive cost calculator

dashboard.html — a dark-themed token cost dashboard that lets you see how much you'd spend using different AI model providers.

compact-dashboard.html — a lightweight version: no charts, 4 top models compared side by side. Each model card shows estimated total cost for your token data at a glance.

Try it live:
Full dashboard — charts, model selector, date range filter
Compact dashboard — 4 models, instant cost comparison

To run locally:

1. Open dashboard.html or compact-dashboard.html in a browser (or serve via any HTTP server)
2. They load paths.json → data.json + models.json
3. Select a model — prices auto-fill from Artificial Analysis data
4. Tweak token counts — costs recalculate instantly

Example files included:

• example-paths.json — points to example-data.json and models.json
• example-data.json — 7 days of synthetic token data for demo

To use your own data, rename example-paths.json → paths.json, point it at your data file, and update your data.json with real token counts.

License

GPL-3.0 — Copyright (C) 2026 Anton Maurer

Credits

• Original scraping concept by demianarc/artificialanalysisscrapper
• Model data source: artificialanalysis.ai