How the price record on this site is built, and the specific things it refuses to claim.
Source
Every figure comes from an Internet Archive capture of a vendor’s own public pricing page. Captures are fetched once and retained locally with their sha256, so a published claim always rests on a specific archived document rather than on a live page that may since have changed. Vendors migrate these pages; the index tracks a lineage of URLs per vendor rather than a single address, because a silently abandoned URL would make the record report “no changes” indefinitely.
Dates are brackets
We never observe the moment a vendor edits a page — only two captures that disagree. So no entry claims a change happened on a date. Each says the change was in place by a date, and states how wide the gap to the previous observation was. A change we could only locate to within more than 45 days is not published as dated at all.
What is refused
A published entry is a dated, sourced claim that a named company changed its prices. A missed change costs a day; a fabricated one is not recoverable. Every check below therefore resolves ambiguity by withholding, and withheld candidates are listed publicly rather than dropped:
- Gaps are not events. — a model absent from a capture is unobserved, never delisted or free.
- Provenance changes are not events. — if the page markup or the source URL changed between the two captures, a price move and a parsing artifact are indistinguishable.
- Renames are not events. — if the vendor’s own label changed — a preview SKU reaching general availability, for instance — a price move and a re-SKU are indistinguishable.
- Unreleased products have no price. — a figure on a card the vendor marked “coming soon” is a placeholder; nothing was purchasable at it, so a later change from it is not a price rise.
- Single-capture blips are not events. — a new price must persist into a later independent capture, and the price it moved away from must itself have held for more than one observation.
- Implausible jumps are withheld. — a move beyond tenfold is almost always a unit misreading rather than a real change.
Model identity
A vendor’s product name maps to a tracked model only through an explicit rule. There is no fuzzy matching, because fuzzy matching is how two different products collapse into one series and a price difference between them is published as a price change. Variants and preview SKUs — which the vendors price separately — are tracked as separate models. A name with no rule is reported as unrecognised rather than guessed at.
Scope and limits
- Published list prices only, for text tokens on standard on-demand endpoints. Negotiated, committed-use, enterprise and batch pricing are out of scope.
- Prices a vendor bands by context length are recorded and labelled per band. A band with no stated upper limit is recorded but not compared across bands.
- Coverage is not uniform. Archive capture frequency varies by vendor and by page, and some historical page layouts are not parsed. Unparsed captures are recorded as gaps, never as zero prices.
- Evidence is a third-party archive copy of a vendor page. That is strong, dated and independently re-checkable, but it is a copy.
Corrections
If a figure here is wrong, the specific archived capture it rests on is linked from the entry. Corrections that change a published figure will be made on the page and noted in the changelog below rather than silently edited.
Changelog
- v1.1 — 03 August 2026. Re-collected: 1,861 captures across 3 vendors and 60 models. The headline changed. v1.0 reported that every change on record was a cut; wider coverage found one genuine increase (Gemini 3.1 Flash Image Preview, input $0.25 → $0.50, by 3 March 2026), so the claim is now stated as a ratio rather than an absolute. The OpenAI o3 cut was also re-dated from 17 to 11 June 2025 as denser captures narrowed its bracket from 8 days to 3 — the figures were always right, the date is now more precise. Two model-naming defects were corrected: preview SKUs and modality/version variants are now tracked as separate models. 10 episodes published, 10 candidates withheld.
- v1.0 — 2 August 2026. First publication. 1,574 captures parsed across 3 vendors and 57 models; 9 episodes published, 8 candidates withheld.