About
Nobody can tell you what API-based AI will cost before you build it. Teams estimate by running a pilot and extrapolating. Finance wants a number before sign-off. Published benchmarks measure model quality, not cost per business task.
This index collects real, measured token costs for specific business tasks so finance and ops teams can put a defensible number on a business case. Each entry records how many tokens a defined task consumed, measured from actual usage, with enough methodology detail to make entries comparable.
Open data, closed write path. Anyone can read, query, and export the dataset. Submissions go through a validated form and a moderation queue.
Quality rules
- Minimum 10 task units measured. Single-run anecdotes are noise. The database enforces this with a check constraint.
- Methodology is required. Contributors describe how they measured tokens, what the inputs looked like, and what tool they used. This is the trust mechanism.
- Median and p90. Median alone hides variance. The 90th percentile is encouraged to show the heavy tail.
- Task unit is explicit. Every entry declares its unit (per invoice, per clause, per ticket). Costs are meaningless without it.
- Self-reported single runs are declined. The moderation queue exists to enforce these rules, not to rubber-stamp submissions.
How moderation works
- A contributor signs in with GitHub (or magic link) and submits an entry through the form.
- The entry lands in the database with
pendingstatus. It is not visible on the site or in the API. - A moderator reviews it in Supabase Studio and sets the status to
approvedorrejected. - Approved entries appear on the site and in the API immediately. The nightly export includes them in the next snapshot.
There is no admin review interface in v1. Moderation happens directly in the database tools. A review queue with diffs is planned for a later version.
Token costs, not dollars
Entries store token counts, not dollar amounts. Prices change monthly. Dollar conversion happens at display time using a pricing file that is refreshed by hand against provider pricing pages. Every dollar figure on the site shows its pricing date so you know how fresh the number is.
Licensing
- Code: MIT Licence.
- Dataset: CC BY 4.0. Attribution keeps the project name travelling with the data.
Get involved
Submit an entryif you have measured token costs from a real project.
Read the API docsto pull the dataset programmatically.