The archive
Anyone can read today's prices off a provider's own page. Almost nobody kept yesterday's. This is what we kept, how complete it is, and how to use it.
165 captures · 2,986 models each · 2024-01-01 to 2026-08-02
What is in it
| Captures | 165 |
| Models in each capture | 2,986 |
| First capture | 2024-01-01 |
| Most recent | 2026-08-02 |
| Span | 945 days |
| Compressed size | 5.0 MB |
Each capture records, for every model a provider publishes a rate for: input price, output price, cache-read price where one exists, context window, and the capability flags the source records. Not a sample. The whole list, every time.
How complete it is
This is the part most archives do not tell you, so it goes near the top rather than in a footnote.
The archive is weekly for most of its span, not daily. 165 captures cover a 945 day period, which is 17.5 percent of days. Captures ran roughly weekly from January 2024 and became daily on 2026-07-29, when the nightly job started. Every figure on this site that cites a date cites a date we actually hold.
What that does and does not support: a weekly series is enough to say what a rate was in a given month and how a workload shape has moved between months, which is what the indices measure. It is not enough to catch a change that was announced and reversed inside a week. Where a rate moved between two captures we know it moved and we do not claim to know the day.
The indices are monthly precisely because the underlying capture is not daily. They chain month to month, and the method is published in full on the methodology page.
Where it comes from
Rates are captured from the LiteLLM public pricing dataset, a continuously maintained community record of provider list prices, and stored in our own repository. We do not scrape provider pages and we do not retype figures by hand.
That means our archive inherits that dataset's coverage and its errors. What we add is time: the dataset publishes what is true now, and replaces it. We keep what it said on each date we captured, which is the thing that cannot be reconstructed later.
Every capture is committed to version control, so the history of the history is auditable too. If a figure here ever changed after publication, the commit that changed it exists.
Check it yourself
The most recent capture has this SHA-256, so a file you download can be verified against what we published:
Spot-check any figure against the provider's own pricing page. If one disagrees, the provider is right and we want to know.
Downloads
| /prices-latest.jsonEvery tracked model at the most recent capture: input, output and cache-read rates. | 5.5 KB |
| /models-history.jsonThe price history the trajectory chart draws, for the tracked basket. | 5.1 KB |
| /movers.jsonThe most recent rate change for each model that has had one. | 2.1 KB |
| /index-series.jsonThe four index series, every reading, with the cached-input share each assumes. | 3.3 KB |
| /index/series.jsonThe living index series, lengthening by one reading a month. | 69.7 KB |
| /index/series.csvThe same series as CSV. | 26.9 KB |
| /workload-cost.jsonWhat each workload shape costs on each model at current rates. | 550 B |
| /estimator-spec-v2.jsonEvery assumption the estimator runs on, versioned. | 28.9 KB |
The raw captures themselves live in the repository under data/snapshots, one gzipped JSON file per date.
Using it
Use the figures, with attribution and a link. If you publish a number from here, cite the dated issue rather than the living page, because a dated issue is frozen and will resolve to the same figures later.
To show a reading rather than quote one, there is a badge: one index value, updated monthly, as an iframe, a script tag or an image.
Corrections and questions: the site is small enough that a real person reads them.