# benchmarks.finance

> Finance-AI benchmark leaderboards and a directory of finance benchmarks, created and maintained by Moneda. Scores are quoted per benchmark with their metric, source and date, and are never combined into one number.

## Pages

- [Leaderboards](https://benchmarks.finance/leaderboards/): every benchmark's results
- [Best by use case](https://benchmarks.finance/best/): the leading models per finance task
- [Directory](https://benchmarks.finance/benchmarks/): the finance benchmarks tracked, with maker, licence and public split
- [Compare](https://benchmarks.finance/compare/): two models on every benchmark where both have a score
- [How Moneda Bench works](https://benchmarks.finance/method/): the six Moneda benchmarks, their tracks, splits and grading
- [About](https://benchmarks.finance/about/)

## Data

- [Full listing](https://benchmarks.finance/llms-full.txt): every benchmark with maker, use case, licence, public split and page URL, and every use case with its current leader (plain text)
- [Search index](https://benchmarks.finance/search.json): every benchmark, model and use case with its page URL (JSON)
- [Changelog feed](https://benchmarks.finance/changelog.xml): when each benchmark, model and score was posted (Atom); page at https://benchmarks.finance/changelog/

## Every page as Markdown

Each content page has a plain Markdown twin at the same path plus `.md`, with the same facts (every score with its metric, source URL and date) and a `Source:` line naming the page. Pattern: `https://benchmarks.finance/leaderboards/<slug>/` is `https://benchmarks.finance/leaderboards/<slug>.md`; likewise `/benchmarks/<slug>.md`, `/best/<use-case>.md`, `/compare/<a>-vs-<b>.md`, `/method.md`, `/about.md`, `/changelog.md`. Index twins:

- https://benchmarks.finance/index.md
- https://benchmarks.finance/leaderboards/index.md
- https://benchmarks.finance/best/index.md
- https://benchmarks.finance/benchmarks/index.md
- https://benchmarks.finance/compare/index.md

## Submitting to us (for agents and scripts)

- Endpoint: POST https://benchmarks.finance/api/submissions with a JSON body
- Schema: GET https://benchmarks.finance/api/submissions returns the JSON Schema of each kind and a guide
- Kinds: correction (a wrong score, date, licence or credit; needs a source URL), model (run a model or product on the benchmarks), benchmark (add a finance benchmark to the directory), product (list a finance AI product with its functionality), contact (anything else)
- Every submission needs a sender name and reply email; a program adds `agent` with its own name and version
- A 202 response returns a reference id; submissions are read by people, not applied automatically
- No authentication; a few submissions a day per sender at most

## WebMCP (agents in the browser)

In browsers that support WebMCP (document.modelContext), every page registers these tools:

- search: models, benchmarks and use cases matching a query, with page URLs
- get_leaderboard: a benchmark's boards, ranked rows with metric, source and date
- get_model_scores: every benchmark score for one model, with rank and source
- best_for_use_case: the leader on each benchmark in a use case; never a combined score
- compare_models: two models side by side on the benchmarks they share
- submit: fills in the contact form; nothing is sent until the visitor presses Send, then it returns the reference id

Rows reported by "moneda" were run by Moneda. A name ending "(public set)" is a run on a benchmark's public questions, not the benchmark owner's result. The data behind the read tools is at https://benchmarks.finance/api/scores.json.

## Rules

- Nothing on the site is investment advice. Moneda is not a regulated financial institution and does not manage portfolios.
- Scores quoted from other benchmarks belong to their authors; each shows its source and date.