Price change log
When model prices moved, and by how much.
Providers change their prices and the old number quietly disappears. This is a dated record of the moves we have observed: what the rate was, what it became, and where the figure came from. Newest first.
Already in effect
| Effective | Model | Input | Output | What happened |
|---|---|---|---|---|
| August 12, 2026 | Claude Sonnet 5 Anthropic | $2 (unchanged) | $10 (unchanged) | The scheduled increase to $3 and $15 on September 1 was cancelled. The introductory $2 and $10 became the standard price, so the rate did not move. Kept as an entry because the announced change was real, was published for weeks, and any forecast built on it is now too high. Read from provider documentation |
| July 30, 2026 | GPT-5.6 Terra OpenAI | $2.50 to $2 (-20%) | $15 to $12 (-20%) | Cut 20% on both input and output, three weeks after the model launched. From rates our catalog was tracking |
| July 30, 2026 | GPT-5.6 Luna OpenAI | $1 to $0.20 (-80%) | $6 to $1.20 (-80%) | Cut 80% on both input and output in the same round, a far deeper cut than its mid-tier sibling. From rates our catalog was tracking |
| July 9, 2026 | GPT-5.6 Terra OpenAI | $2.50 | $15 | Launched at $2.50 input and $15 output, rates that held for three weeks. From rates our catalog was tracking |
Why this page exists
A price change gets written about on the day and is then almost impossible to look up. Rate cards show today’s number, announcements scroll away, and third-party pricing tables are overwritten in place with no record that anything moved. If you are trying to explain last quarter’s bill, or to forecast the next one, the history matters more than the current figure.
Keeping this is a side effect of how Spendline works rather than an editorial project. The product compares every model’s rate against the providers on a daily run, because it has to price real API traffic correctly. That job sees a provider’s old and new price side by side, which is exactly what a change log is made of.
Two things this page is not. It is not complete: it starts where our records start, and it covers the models we track rather than every model in existence. And it is not automatic guesswork. Entries marked as read from publisher documentation were checked against the provider’s own pricing page; entries marked as observed came from dated rates our catalog was already carrying. Where we cannot point at a source, there is no entry.
What the moves so far suggest
Cuts have been deeper at the cheap end than the expensive end. In the same round on July 30, 2026, OpenAI took its small model down 80% and its mid-tier model down 20%, while the flagship did not move. Anthropic ran the other way and then reversed: the introductory rate on Claude Sonnet 5 was published as expiring into a 50% rise on September 1, 2026, and on August 12 that increase was cancelled and the lower rate made permanent. Both directions carry the same lesson. A model tier is not a stable price point, an announced date is not a settled one, and a budget built on one month’s rate card is a budget with an expiry date on it.
A rate change you did not notice is a margin change you cannot explain
Knowing the rate moved is the easy half. The hard half is knowing which customers and which agents were on that model, and what it did to their cost. Spendline records every call against the customer it belongs to, so a price change becomes a number you can point at rather than a surprise at month end.
Run the 5 minute assessment