Attribution and licences
We aggregate published data and we do not run our own evaluations, so everything on this site belongs to someone else first. This page names every publisher, the licence their results reach you under, and exactly which evaluations came from each. It is generated from the database on every build, so it cannot drift from what we actually redistribute.
CC-BY 4.0 · credit required · 6 evaluations published directly · 7 more reaching us through their redistribution
Epoch AI, 'AI Benchmarking Hub'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks' [online resource].
Source: https://epoch.ai/benchmarks · Licence: CC BY 4.0. This credit covers every row we take from Epoch AI, not only the evaluations Epoch AI authored. 7 further evaluations — authored by LiveBench and listed below — reach us through the same CC-BY 4.0 benchmark archive at epoch.ai/benchmarks, so Epoch AI's citation is required for those too. Credit both: the author of the benchmark, and Epoch AI for the redistribution we actually obtained it from. We have not modified the published scores; where they feed an index they are min-max normalised alongside every other model's, and the original values remain in the score download.Published by Epoch AI:
Redistributed through Epoch AI under CC-BY 4.0, authored by LiveBench — Epoch AI's credit applies to these as well:
CC-BY 4.0 · credit required
| GPQA Diamond · index component | 169 models |
| Humanity's Last Exam · index component | 45 models |
| MATH / AIME · index component | 142 models |
| MATH Level 5 | 101 models |
| SWE-bench Verified · index component | 32 models |
| Terminal-Bench · index component | 55 models |
Apache-2.0 · redistributed via Epoch AI
LiveBench authored these evaluations; we obtain the scores from Epoch AI's CC-BY 4.0 benchmark archive (epoch.ai/benchmarks), so that credit is required too and is given above alongside LiveBench's own.
| LiveBench · index component | 52 models |
| LiveBench Coding | 52 models |
| LiveBench Data Analysis | 52 models |
| LiveBench Instruction Following | 52 models |
| LiveBench Language | 52 models |
| LiveBench Mathematics | 52 models |
| LiveBench Reasoning | 52 models |
Prices, throughput, latency, context windows and model metadata are ingested from these sources under their own terms. Elo ratings come from blind human preference votes we did not run and do not adjust. Vendor logos are the trademarks of their owners, self-hosted here and used only to identify the model or provider they belong to.
What we add, and how you may use it
Our contribution is the compilation: the identity matching that decides two names are the same model, the normalisation, the Known Good Index and the capability indices, and the pages around them. That compilation is free to use with credit to The Known Good and a link to theknowngood.com.
The underlying data stays under its own licence. A permission we grant cannot extend past what the original publisher granted us. One of the publishers above — Epoch AI — requires credit, and that requirement follows the data: if you redistribute those figures, from this site or from anywhere else, you must carry the credit yourself, exactly as it appears above. Where a benchmark reaches us through someone else's archive, both are owed: the author of the evaluation, and the publisher who redistributed it. When in doubt, credit the original publisher rather than us.
The same figures are downloadable as CSV and JSON, with a licence note in the JSON payload pointing back to this page. No account, no key, no rate limit.
Corrections
If a number here misrepresents a publisher's result, that is our error and we want to fix it before you have to argue with it. Every score links to the evaluation page naming its publisher, and every evaluation page links to the publisher's own page — start there, then tell us. Corrections take priority over everything else we are building.