# LLM Compare — Machine-Readable Data

> A comparison tool for LLM benchmark scores. All data is static JSON served from GitHub Pages.

## How to Use This Site's Data

This site displays benchmark data from [ZeroEval](https://api.zeroeval.com) and arena scores from [Magia](https://magia.land). All data is pre-fetched and committed as static JSON files, so you can access it directly without executing JavaScript.

## Data Endpoints

### Model List
```
GET https://broskees.github.io/llm-compare/data/models.json
```
Returns a JSON array of all tracked models. Each entry includes `model_id`, summary benchmark scores, and metadata.

### Model Details (Benchmarks)
```
GET https://broskees.github.io/llm-compare/data/details/{model-id}.json
```
Returns full benchmark breakdown for a specific model, including individual task scores.

### Arena Scores
```
GET https://broskees.github.io/llm-compare/data/arena/{model-id}.json
```
Returns Magia arena scores for categories: chat-arena, text-to-text, text-to-website, text-to-game, p5-animation, threejs, dataviz, tonejs. Returns `{}` if no arena data exists.

## URL Patterns

### Comparison URLs
The site uses `?m=` with comma-separated model IDs:
```
https://broskees.github.io/llm-compare/?m=claude-opus-4-6,gpt-5.3-codex,kimi-k2.5
```

### Resolving a Comparison URL to Data
1. Parse the `m` query parameter (comma-separated model IDs)
2. For each model ID, fetch:
   - `data/details/{model-id}.json` for benchmarks
   - `data/arena/{model-id}.json` for arena scores
3. Use `data/models.json` to get the full model list and metadata

### Model ID Format
Model IDs use lowercase with hyphens. Slashes in original IDs are replaced with `--`. Examples:
- `claude-opus-4-6`
- `gpt-5.3-codex`
- `kimi-k2.5`
- `deepseek-r1-0528`

## Data Sources
- Benchmarks: [ZeroEval API](https://api.zeroeval.com)
- Arena Scores: [Magia API](https://api.zeroeval.com/magia/models/scores)
- Data is refreshed periodically via GitHub Actions

## Repository
https://github.com/broskees/llm-compare