What's new
Analysis14 Nov 20265 min readData as of 14 Nov 2026

Our price history has 9,155 later records. 9,146 repeat the price before.

A price history should show when prices moved. Ours mostly records the same price again. The few recorded differences in it cannot yet be told apart from relabelling and corrections.

GPT-5.5 has 110 rows in our price history between 25 July 2026 and 14 November 2026. All 110 say the same thing: $5 in, $30 out. OpenAI's own model page, read on 14 November 2026, lists the same prices, so the price itself is right. It is just a lot of rows for a recorded price that never changed.

The pattern is general. Across 107 language models with prices from sources other than OpenRouter, we hold 9,262 rows. Take the first row of each model as the starting point, and 9,155 remain. Of those, 9,146 (99.9%) have exactly the same input and output prices as the row before. 9 do not.

Why the history fills up with repeats

The rows come from two paths. 3,215 are labelled with a lab's name and "(official)", the labels used in the catalogue that our official-pricing importer copies from. That script copies prices typed into the catalogue by hand and, for every model in its list that already has a database record, writes the price, stamps the "pricing updated" time and adds a row to the price history, whether or not the price differs from the last run. We checked this in the script's code on 14 November 2026. It makes no request to a lab's website. The only trigger for a new row is that the script ran.

The other 6,047 rows are labelled with a website address, such as a lab's pricing page. We have not traced which collector writes those. They show the same pattern: 5,965 of 5,966 later rows repeat the price before, against 3,181 of 3,189 for the "(official)" rows.

So an "(official)" row tells you the importer ran. It does not tell you that anyone looked at the lab's page that day, and the "pricing updated" time that importer writes is no better evidence. We do not know what triggers the other rows. Our transparency page lists this as a known fault.

The 9 rows that changed

9 later rows differ from the one before. Each is listed below, with the source label on each side.

Source: AI Resource Hub price history for language models, rows not sourced from OpenRouter, cache of 14 November 2026. US dollars per million tokens. "Label" is the source name recorded on the row.
ModelDateBeforeAfterLabel changed?
Claude Opus 4.615 March 2026$5 / $25$15 / $75Yes
Claude Opus 4.625 July 2026$15 / $75$5 / $25No
DeepSeek V315 March 2026$0.14 / $0.28$0.27 / $1.1Yes
DeepSeek V3 032415 March 2026$0.14 / $0.28$0.27 / $1.1Yes
Mistral Large15 March 2026$0.5 / $1.5$2 / $6Yes
Mistral Nemo15 March 2026$0.02 / $0.04$0.15 / $0.15Yes
Mistral Small 3.1 24B15 March 2026$0.03 / $0.11$0.1 / $0.3Yes
o323 February 2026$2 / $8$10 / $40No
o315 March 2026$10 / $40$2 / $8Yes

7 of the 9 came with a change in the source label, all of them on 15 March 2026, when the name recorded on the row switched from a website address to a lab's name. A price that moves at the same moment as its label could be a real change, a correction to an earlier mistake or the result of switching from one reading of the lab's page to another. The history cannot say which.

The other 2 keep the same label. Claude Opus 4.6 moves from $15 / $75 to $5 / $25 in July, and o3 moves from $2 / $8 to $10 / $40 in February. We have not checked either against the lab's page at the time, so we cannot say whether they are corrections or real moves, and we have not counted them as cuts or rises.

Why this blocks a "how often do prices change" story

A claim such as "most models never get cheaper" is tempting and easy to produce from a table like this one. It would also be hard to defend. 99.9% of the rows are repeats, the handful of moves cannot be separated from relabelling, and the rows before and after the relabelling carry different source names. Counting changes in it would measure how the data was kept, not how prices behave.

That is why we have not published a price-cut article. We think a usable one needs a single source per model, a "last checked" date that only a person or a real fetch can set, and the lab's effective date kept apart from both. Until then, the fair reading of our history is that it records what our catalogue said on each run.