Markdown & JSON API

1. Find an identifier

Start with a directory. Copy the model slug or benchmark id exactly as returned, then append .md or .json to its detail URL.

curl -fsS 'https://benchize.com/api/v1/models?search=GLM&limit=5'
curl -fsS 'https://benchize.com/api/v1/benchmarks?search=GPQA&limit=5'

Use a model’s slug for one configuration, such as -thinking or -high. An unspecified configuration has no suffix. A benchmark id may contain a colon: URL-encode it instead of changing it.

2. Choose Markdown or JSON

Both formats contain the same published reports and original sources. Detail documents have no separate HTML page; browser directory links open a result dialog.

Markdown

Readable reports with tables, evidence links and a citation footer. Share the URL with an AI assistant or read it directly.

JSON

Structured reports for scripts and integrations. Metrics, configuration, statistics and provenance retain their original data types.

HTML directories also accept the Accept: text/markdown header. Explicit .md links are the easiest format to share and cache.

curl -fsS -H 'Accept: text/markdown' 'https://benchize.com/models'

3. Search and follow pagination

Use the public list APIs to find identifiers. Add the same search or developer/category filters when you follow a cursor.

A detail response is one page, not necessarily all results. JSON includes nextCursor and a Link header with rel="next"; Markdown includes a continuation link. Keep the original filters on each subsequent request.

Example: read every report

const root = 'https://benchize.com';

async function readReports(collection, id) {
  const path = '/' + collection + '/' + encodeURIComponent(id) + '.json';
  const url = new URL(path, root);
  const reports = [];
  let nextCursor;

  do {
    const response = await fetch(url);
    if (!response.ok) throw new Error('HTTP ' + response.status);
    const page = await response.json();
    reports.push(...page.evaluations);
    nextCursor = page.nextCursor;
    if (nextCursor) url.searchParams.set('cursor', nextCursor);
  } while (nextCursor);

  return reports;
}

4. Understand and cite the data

Benchize collects public evaluation reports. It does not run inference or calculate a new composite score. Keep each report’s original source citation.

Compare reports only when the benchmark, publisher, sourceVersion, metric, unit, scope and direction match. Preserve uncertainty and configuration; missing values are not zero.

Caching and freshness

Explicit Markdown, JSON and llms resources may be cached at the edge for up to 24 hours. New data can take that long to appear in a cached document; a code release switches to a new cache version.

Reuse the ETag value in If-None-Match. A 304 response has no body: keep your existing copy. X-Benchize-Document-Cache reports HIT, MISS or BYPASS for this document cache.

When you need fresh data immediately, send Cache-Control: no-cache. Ordinary reads can reuse the cached response.

curl -i -H 'Cache-Control: no-cache' 'https://benchize.com/models.md'

HTTP status and errors

Use with an AI assistant

Give your assistant the main llms index for usage instructions, or a complete section index to discover every eligible document. llms-full.txt links to bounded content shards.

• Usage and selected entries /llms.txt
• Complete model index /llms/models.md
• Complete benchmark index /llms/benchmarks.md
• Full-content shards /llms-full.txt