About

We run no benchmarks

Not one figure here was produced by us. Every score is one a model's maker published in a model card, paper or announcement, or one an independent platform measured and published — and every score links that document. If a figure cannot be traced to a source, it is not shown; the cell stays empty.

That is a limitation as much as a principle. It means the table can only be as complete as what people publish, and it means a lab that reports nothing looks the same here as one that reports badly. The upside is that anything you read can be checked, by you, in one click.

How to read what is here

• Two kinds of figure, kept apart. What a lab published about its own model is a different claim from what an independent platform measured. Both appear; the measured ones are in their own colour.
• An empty cell means unpublished, not zero. A sparse row is a claim never made rather than a claim that failed.
• Effort is part of a score. Where a source says how hard the model was made to think, the figure carries it. Where it does not, the tooltip says so instead of implying a default.
• Compare down a column, not across the table. Labs evaluate on their own harnesses; two of them reporting the same benchmark have not necessarily run the same test.
• Prices are list rates and go stale. Vendor pay-as-you-go pricing as of the date shown, excluding batch, cached-input, off-peak and priority tiers.

The longer version is on Benchmarks, which explains what each one measures and how to run it yourself.

Using the data

The catalog is published as JSON and is free to use with attribution to this site. The benchmark figures themselves belong to whoever published them — each score carries its own source_url, so anything you build on can be traced back past us to the original.

There is also an llms.txt for agents that arrive without a browser, and an RSS feed for the writing.