The Known Good Updated 25 Jul 2026

Download the data

Every figure behind every chart, as flat files. Free, no account, no key, no rate limit — the point of the site is that you can check our working, which is difficult if you cannot have the numbers.

Rebuilt with the site, hourly UTF-8, header row, one row per record Missing values are empty, never zero
Files
models.csv
908 rows · one per tracked model
Slug, name, creator, release date, Known Good Index and its component count, coding and openness indices, input / output / blended price, output tokens per second, time to first token, context window, max output tokens, reasoning and open-weights flags, licence, status.
CSV
eval_scores.csv
908 rows · one per model per evaluation
Evaluation name, publisher, licence, whether it feeds the index, model, the published score as published, and the date it was measured. This is the file to check us against — the scores are unmodified.
CSVPublisher licences apply
arena.csv
1,025 rows · one per model per arena
Arena kind, model, creator, Elo, the 95% confidence bounds, vote count and published rank. Ingested from Arena, and from TTS Arena for speech.
CSV
models.json
908 models · nested
The same models with every index, every evaluation score, price, performance and capability metadata nested per model, plus the build timestamp and a licence note pointing back to attribution.
JSON

Licence

The compilation is ours; the measurements are not. Use our compiled figures — the identity matching, the normalisation, the Known Good Index — freely, with credit to The Known Good and a link to theknowngood.com.

The underlying results stay under their publishers' licences, and some of those require credit. Epoch AI benchmark data is CC-BY: if you redistribute it, from these files or anywhere else, you must carry Epoch's citation. It is written out in full, alongside every other publisher and licence we hold, on the attribution page — read that before you republish eval_scores.csv.

Notes for anyone parsing this

  • Empty means unknown. An empty cell is a figure we do not hold. It is never a zero, and we never interpolate one. Treat empty and 0 as different values.
  • Known Good Index is 0–100 and ours. It is not any publisher's headline scale, and a row with no index value is a model missing science, mathematics or coding evidence among the component evaluations — methodology explains why we would rather publish nothing than a score built on one area alone.
  • Scores are relative to what we track. Normalisation is against the tracked population, so index values shift as models are added. Take the build timestamp in models.json with you.
  • Prices and speeds are snapshots of the latest ingest, not averages over time, and speed varies by host.
  • These files are the stable interface. The other JSON under /data/ exists to drive the charts on this site, changes shape whenever a page changes, and is not an API. Please use the four files above.

Corrections

If a figure in these files disagrees with the publisher it came from, the publisher is right and we want to know. Start from the evaluation pages, which name the publisher for every score, and see contact.

Licence — you may use these exports commercially, with credit

Our own contribution — the compilation, the normalisation, and our derived measures including the Known Good Index — is published under the Creative Commons Attribution 4.0 International licence (CC BY 4.0), per Terms of Use section 4.3. You may download in bulk and use these files for research, journalism, analysis, teaching and commercial purposes. No permission needed, no fee.

Attribution runs to two parties. Credit The Known Good with a link to https://theknowngood.com — and retain the upstream publishers' own attribution, which passes through unchanged and which our licence cannot waive. Epoch AI data is CC-BY and requires credit. Every source and its licence is listed on attribution.

This licence is also embedded inside every file — as a comment block above the header row in each CSV, and as a license object in the JSON — so it travels with the download.

The Known Good
The Known Good aggregates publicly published benchmark data. We do not run our own evaluations. Data under respective source licences. Full source list: attribution · How each number is produced: methodology · Last run of every source: about.