LLM Price Series

Model prices change without announcement, models are deprecated, retired, sunset and withdrawn without notice, and no provider publishes a history of deprecations or price changes. The data is public every day and archived by nobody.

The questions this makes answerable have no other source. What did this model cost in March, when we quoted that contract. Which models have been withdrawn since we shipped, and were we calling them. Did our unit economics move because we changed something, or because a price did.

YOU ARE NOT BUYING TODAY'S PRICES. Those are free from the source. You are buying the days you did not collect -- and no budget closes that gap, which is the shape of every accumulation dataset: cheap to hold, impossible to backfill, worthless if you start next year.

TWO RULES SEPARATE A SERIES FROM A PILE OF FILES.

First, A FAILED COLLECTION IS NOT A MARKET EVENT. A fetch returning 200 with an empty body looks exactly like every model being withdrawn on the same day. Record that once and the dataset answers its most valuable question with fiction, and nothing about the file says so. A failed fetch writes NOTHING and exits non-zero: a missing day is an honest gap, an empty file is a lie. The diff refuses outright when the newer snapshot is empty and the older is not.

Second, A WITHDRAWAL PRODUCES NO ROW. A diff that iterates the new list structurally cannot see a model that is gone, because absence contributes nothing to iterate over. Miss it and the series records every arrival and no departure -- looking complete while being wrong about the thing people pay for. Withdrawals are found by walking the OLD set.

PRICES COMPARE AS DECIMAL STRINGS, NEVER FLOATS. The source publishes values like 0.00000009, and round-tripping that through parseFloat introduces differences at the last digit which present as price changes nobody made. A fabricated price change is worse than a missed one: it is an event someone will investigate.

WHAT SHIPS: the full collector and diff engine, every snapshot collected to date as plain JSON with no database required, daily change files so the interesting part is readable without replaying the series, and a launchd plist for unattended daily collection. 338 models tracked across 13 pricing dimensions -- prompt, completion, cache read and write, audio, image, web search, internal reasoning. Re-running the collector is safe: a day already collected is never overwritten, because silently replacing yesterday's file with today's prices is a corruption nothing would report.

WHAT THIS DOES NOT CLAIM: it records PUBLISHED prices from a public aggregator. It does not include private or negotiated rates, and it cannot tell you what a provider charged before collection began. There is no way to buy that; there never will be. That is the entire point.

VERIFIED: 29 tests, 10/10 mutations caught. Collecting daily since 2026-08-04. Zero runtime dependencies.

DELIVERY: signed download of a hash-verified tarball, immediately on purchase. Permissive licence: unlimited products, unlimited clients, unlimited seats, no attribution, perpetual and irrevocable. One restriction, do not republish the source as source.

Interface

What you call, and what comes back. Types and signatures only — the implementation ships with the source.

  export async function collectOnce(io: CollectorIo = realIo, base: string = BASE): Promise<CollectionOutcome>;
  export function normalisePrice(v: string | number | null | undefined): string | null;
  export function percentChange(before: string | null, after: string | null): number | null;
  export function diffSnapshots(before: Snapshot, after: Snapshot): Change[];
  export function parseModelList(payload: unknown, takenAt: string, source: string): Snapshot;
  export function summarise(snapshots: Snapshot[], changes: Change[]): SeriesSummary;

  export type ChangeKind = "added" | "withdrawn" | "price_changed" | "context_changed";