Methodology

What this is

A directory of 40 LLM benchmark leaderboards, checked on 2026-10-31. For each one it answers four questions people ask before they build on leaderboard data: where are the raw results, may I reuse them, can I submit a model, and is the board still maintained? It does not copy any scores. Open the leaderboard itself for those.

This is not legal advice. Licences and terms change, and our verdict is a short summary of the quoted text. Read the licence yourself before you republish anything, and ask the operator when in doubt.

Where the data comes from

• Raw data: the leaderboard page, its GitHub repository or Hugging Face dataset/Space, and its API or download docs. "Not published" means we found no public file, dataset or API for the results.
• Licence: the licence file or licence field of the repository or dataset that holds the results, or the operator's terms or data page. Licence and terms statements are quoted verbatim with a link. A script checks that each quote still appears in the saved copy of the page. Where we only have GitHub's or Hugging Face's licence metadata, we say so.
• Submissions: what the leaderboard, README or submission guide says about outside results, quoted where it says anything. "Not stated" means we found no statement.
• Last updated: the newest of the last commit to the results repository or file (GitHub API), the Hugging Face dataset or Space lastModified, or a date stated on the page. Over 90 days old counts as "no update in 90+ days". A board without any readable date is shown as "date not published".

Code licence or data licence?

Many leaderboards keep their results inside a code repository. When the repository licence (for example MIT or Apache-2.0) covers the very files that hold the results, we count it as "reuse with attribution". When the results sit somewhere else, or the licence clearly targets the code or app only, we mark it "unclear" and explain why. Aggregator sites often say that the underlying benchmark results keep their publishers' own terms; we quote that too.

The labels

Licence / reuse

• Public domain (CC0): Released under CC0: reuse without conditions.
• Reuse with attribution: A licence or terms allow reuse if you credit the source (for example CC BY 4.0, MIT, Apache-2.0).
• Non-commercial only: Reuse is allowed for non-commercial purposes only (for example CC BY-NC 4.0).
• Unclear: A licence exists, but it covers the code or website rather than clearly the results.
• No licence found: We found no licence or terms covering the data. Link to it, and ask before republishing.
• Reuse restricted by terms: The operator's terms restrict redistribution or competing uses of the data.

Submissions

• Open: Anyone can submit results, for example by pull request.
• Restricted: Submissions only from some groups, by contact, or tested by the operator.
• Closed: The operator says submissions are closed or the board is retired.
• Not stated: We found no statement about outside submissions.

Freshness

• Updated in last 90 days: The newest date we could read from the source is at most 90 days old.
• No update in 90+ days: The newest date we could read from the source is more than 90 days old.
• Archived: The source repository is archived (read-only).
• Retired: The operator has announced the leaderboard is retired.
• Date not published: We found no machine-readable or stated update date.

How results are made

• Operator runs the evals
• Outside submissions
• Crowd votes
• Usage data
• Aggregator of other boards

Left out

We list a leaderboard only when we could confirm where it lives and read its licence situation. Not included yet:

• WebArena: The leaderboard is linked from the README, but we could not confirm where it lives or how it is updated.
• OpenAI simple-evals: An evaluation library with reference numbers, not a maintained leaderboard.
• Vellum LLM Leaderboard: A marketing page; the terms we found cover Vellum's product, not the leaderboard, and no sources or data are published.
• PublicAI Index: A meta-index that rescales other boards; we found no data licence statement on the pages we checked.
• LLMBenchmark.ai: A meta-index built on other leaderboards; no data licence found.
• peerbench DACH: A German-language aggregate ranking; operator and data licence could not be confirmed from its pages.
• OpenCompass academic leaderboard: Not fetched successfully; the OpenVLM leaderboard from the same team is listed.

Missing a leaderboard or found a mistake? Use the bug link in the footer.

Use the data

/api/leaderboards.json (this directory's data under CC BY 4.0; each leaderboard's own data keeps its own licence), llms.txt and the changelog (RSS). Back to the table.