Bertrande

Download the data

Everything on this site is downloadable, in JSON and CSV, free, with attribution. Machines are a first-class audience here, not an afterthought.

Full dataset

FileContentsFormat
current.jsonEvery published parameter at its latest observed value, with date, source and source level. 14586 recordsJSON
current.csvThe same, as a spreadsheetCSV
changes.jsonEvery confirmed change, with before, after and delta. 336 changesJSON
feed.xmlThe same changes as an Atom feedAtom
prices.jsonThe price matrix: both prices, the cheapest tier and its name, currency, observation dateJSON
reliability.jsonAvailability of each upstream host over the previous day, as OpenRouter's router observed it. One reading a day, every reading keptJSON
latency.jsonTime to first token and output throughput, as the Hugging Face router measured them. The median of our own daily readings, with how many and over what periodJSON

Per service

Every one of the 1289 service pages carries its own data.json and data.csv at the same address. The JSON includes the full observation history of each parameter, which is the part of this dataset that cannot be reconstructed after the fact.

Example: .../anthropic--claude-opus-4/data.json

Licence and how to cite

Published under CC-BY-4.0. Reuse is permitted, including commercially, with attribution and a link back. Cite the service page URL and the observation date shown next to the figure. A figure of ours without a date attached is not ours.

What you will not find in these files

Anything sourced only from a third-party aggregator. Those figures exist in our database and are deliberately excluded from every export, for the same reason they are excluded from the pages. See the source ranking.

Speed, and what our figures are worth

Price tables leave speed out, and speed is what developers complain about. We publish time to first token and output throughput, measured by the Hugging Face router on the traffic it sends, and availability, measured by OpenRouter's router on the traffic it sends. Neither is our own measurement, both say so on every page, and neither is what you would get calling a host directly: the network path and the load are not the same.

We never publish the latest reading on its own. A single latency measurement covers a handful of calls. We read once a day, keep every reading, and show the median of ours with how many it rests on and over what period. That median is a derivation, the computation is ours and the measurements are not, and it is calculated on dated observations we do not overwrite. This is the part nobody else offers: the firms that measure speed cap their public history at ninety days or seventy-two hours. Ask us in 2030 what a host was doing in 2026 and we will still have it.

What we refuse: no average of what users report, because an average of anecdotes is a figure nobody measured; and no single speed per seller, because one model can sit behind more than thirty hosts and picking the worst, the best or a mean of them would be a number nobody measured either.

An API is coming

A versioned read API with a generous free tier is planned. These static files are the honest version of it in the meantime: same data, same licence, no key required.