How this index tracks AI prices, and what it deliberately leaves out

This site is a pipeline that happens to render pages. The dataset is the product; the pages are a view of it. That choice has consequences for how you should read every number here, and this article is the plain-language version of the methodology page.

What the dataset contains

• 1,657 endpoints from 32 providers, of which 1,557 chat models publish both an input and an output rate.
• Normalised units. Every rate is expressed in US dollars per one million tokens, so two providers' tables can be compared without a spreadsheet.
• Provenance on every record. Each model carries the sources it was read from, and a second source where one exists.
• Capability flags (vision, tools, reasoning and similar) taken from the same upstream records, never inferred.

The four steps

1. Collection. Rates are read from provider pricing pages and community registries, then normalised into one schema.
2. Cross-check. Where a second source exists, the figures are compared. A disagreement above a couple of percent is shown on the page with both numbers rather than averaged away, because an averaged price is a made-up price.
3. Diff. Each refresh is compared against the previous snapshot. Price moves, additions and retirements become changelog entries with old and new values.
4. Publication. The dataset is published as JSON and as an open mirror, then rendered into the pages here — and, where editorial judgement adds something the raw table cannot, into the articles in this section.

What is deliberately excluded

• Batch and asynchronous discounts — they move real bills and cannot be represented in a single comparable rate.
• Committed-use agreements — per-account, therefore unrepresentable in a public dataset.
• Regional rate cards and free tiers — frequently different from the USD list rate.
• Retries, evaluation and non-token costs — real, but not token pricing.

Curation is part of the method, not a growth hack

One model is often resold by a dozen platforms. Publishing a page per listing would multiply this site by roughly six and add nothing a reader can use, so listings are grouped and each model gets one canonical page that shows every platform's price for it. The duplicate listings stay reachable from that page but are not offered to search engines — the About page explains the policy, and `apps/web/src/lib/indexing.ts` in the public repository is the implementation.

Corrections

If a rate here is wrong, the correction goes into the pipeline rather than onto the page: fix the source mapping, re-run, and the fix propagates everywhere the number appears. Tell us via the contact page — corrections are usually live within one refresh cycle.