Across 60 models from 3 vendors, tracked over 25 months of dated Internet Archive captures of the vendors’ own pricing pages, published API list prices changed in only 10 episodes. 51 of 60 models never changed price at all between their first and last observation.
And they move almost entirely in one direction: 17 of the 18 changes on record are cuts. The sole increase in 25 months is Gemini 3.1 Flash Image Preview — a preview model whose input price went $0.25 → $0.5 by 3 March 2026. The most recent change of any kind was 3 March 2026.
Why this record exists
Vendors replace their pricing pages rather than version them, so old prices simply stop existing. Both pricing URLs this index started from have already been abandoned and redirected elsewhere. What a model cost in 2024 is therefore not retrievable from the vendor at all — only from an archive, and only if someone extracted it before the page moved.
How to read a date
We observe archived captures of a pricing page, not the moment a vendor edits it. Each entry is therefore bounded by two observations: the last capture showing the old price, and the first showing the new one. “By 9 October” means the change was in place by then, and was not in place at the previous observation. Every entry states the width of that bracket, and links both captures so the claim can be checked independently.
Every change on record (10 episodes)
An episode is one vendor moving one model on one date, however many price components moved with it.
By 3 March 2026 — Google raised Gemini 3.1 Flash Image Preview
- input: $0.25 → $0.5 per million tokens +100%
By 24 October 2025 — Google cut Gemini 2.5 Flash-Lite
- cache read: $0.025 → $0.01 per million tokens −60%
By 24 October 2025 — Google cut Gemini 2.5 Flash-Lite Preview
- cache read: $0.025 → $0.01 per million tokens −60%
By 16 October 2025 — Google cut Gemini 2.5 Flash Preview
- cache read: $0.0375 → $0.03 per million tokens −20%
By 9 October 2025 — Google cut Gemini 2.5 Flash
- cache read: $0.075 → $0.03 per million tokens −60%
By 9 October 2025 — Google cut Gemini 2.5 Flash Preview
- cache read: $0.075 → $0.0375 per million tokens −50%
By 9 October 2025 — Google cut Gemini 2.5 Pro
- cache read: $0.31 → $0.125 per million tokens (prompts up to 200k tokens) −60%
- cache read: $0.625 → $0.25 per million tokens (prompts over 200k tokens) −60%
By 11 June 2025 — OpenAI cut o3
- cache read: $2.5 → $0.5 per million tokens −80%
- input: $10 → $2 per million tokens −80%
- output: $40 → $8 per million tokens −80%
By 3 December 2024 — Anthropic cut Claude 3.5 Haiku
- cache read: $0.1 → $0.08 per million tokens −20%
- cache write: $1.25 → $1 per million tokens −20%
- input: $1 → $0.8 per million tokens −20%
- output: $5 → $4 per million tokens −20%
By 4 October 2024 — Google cut Gemini 1.5 Pro
- cache read: $0.875 → $0.3125 per million tokens (prompts up to 128k tokens) −64%
- input: $3.5 → $1.25 per million tokens (prompts up to 128k tokens) −64%
- output: $10.5 → $5 per million tokens (prompts up to 128k tokens) −52%
Held for review (10)
Candidate changes the evidence does not support publishing. They are listed rather than discarded, because silently dropping them would misstate how much this record covers.
| Model | Component | From | To | Held because |
|---|---|---|---|---|
| gemini-robotics-er-preview | input | $0.3 | $1 | the vendor's own label changed between the two captures ('Gemini Robotics-ER 1.5 Preview' -> 'Gemini Robotics-ER 1.6 Preview'); a price move and a rename/re-SKU are indistinguishable here |
| gemini-robotics-er-preview | output | $2.5 | $5 | the vendor's own label changed between the two captures ('Gemini Robotics-ER 1.5 Preview' -> 'Gemini Robotics-ER 1.6 Preview'); a price move and a rename/re-SKU are indistinguishable here |
| Gemini 2.5 Flash Preview | cache read | $0.0375 | $0.075 | bracket is 102 days wide (limit 45); the change cannot be dated with useful precision and the gap may conceal more than one move |
| Gemini 2.5 Flash Preview | input | $0.15 | $0.3 | bracket is 102 days wide (limit 45); the change cannot be dated with useful precision and the gap may conceal more than one move |
| Claude 3.5 Haiku | cache read | $0.025 | $0.1 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
| Claude 3.5 Haiku | cache write | $0.3125 | $1.25 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
| Claude 3.5 Haiku | input | $0.25 | $1 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
| Claude 3.5 Haiku | output | $1.25 | $5 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
| Claude 3.5 Haiku | cache read | $0.03 | $0.025 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
| Claude 3.5 Haiku | cache write | $0.3 | $0.3125 | one side is pre-release placeholder pricing (the vendor's own page marked the model unreleased); no purchasable price changed |
Scope. These are published list prices for text tokens on standard, on-demand API endpoints — not negotiated rates, committed-use discounts, enterprise agreements, or batch-mode pricing. Prices that a vendor bands by context length are labelled with their band. A model absent from a capture is treated as unobserved, never as removed or free.
Data
The full observation set behind this page is available as CSV: ai-model-price-observations.csv. Every row carries the archived capture it came from and the sha256 of that capture, so any figure here can be traced to a specific archived document. Read the methodology.