Every LLM API price change we can date

What this data is, and isn't

Every row below comes from a secondary source: the git history of BerriAI/litellm's public price table, sampled once a month. One snapshot per calendar month: the last commit of that month in BerriAI/litellm, read from raw.githubusercontent.com at that exact SHA, which anyone can fetch and check. A change is recorded on the date of the snapshot that first showed the new number, so the resolution is one month and the date is when the price was first *recorded*, not when the seller announced it. Only entries with a two-sided per-token price are included; embeddings, reranking, image and audio pricing are out of scope. Prices are converted from per-token to per 1,000,000 tokens and otherwise not touched. Two kinds of difference are NOT counted as price changes, because they are the aggregator correcting its own file rather than a seller changing a tariff: anything with a zero on one side (LiteLLM writes 0 both for "free" and for "not filled in yet"), and any jump of tenfold or more inside a single month. Both are kept, with the reason, in excluded_changes -- the rule is stated so a reader who disagrees with it can apply their own.

A date here means "first recorded on", not "announced on". The resolution is one calendar month, so a price that changed and changed back inside the same month never appears, and a change dated the last day of a month may really have taken effect any time since the previous snapshot. This is a different, coarser kind of evidence than the individually re-verified, seller's-own-page dates used everywhere else on this site -- including on the change log for this site's own dataset, a much smaller set of models tracked at seller-page precision rather than monthly aggregator precision.

We separately measured this aggregator against our own primary dataset -- prices read directly off each seller's own pricing page, not off anyone's git history. Of 118 models comparable between the two, 115 were identical (97.5%). Agreement is reassuring, not proof -- both figures could in principle trace back to the same seller announcement. The stricter record, sourced entirely from sellers' own pages and archive snapshots of them, is what tokens used to cost; use this page for reach across the long tail, that one when a single number has to be unimpeachable.

What this shows

Of 496 recorded changes, 286 were price cuts and 210 were price rises. 1240 model entries were delisted at some point in the window and never came back.

The steepest recorded output-price cut is openrouter/mistralai/mistral-small-3.1-24b-instruct (openrouter): $0.30 to $0.027 per 1M output tokens, first recorded July 31, 2025 -- a drop of -91%. (Moves that land at exactly $0 are excluded from this ranking; they usually mark a model made free, not a comparable price cut.)

openrouter shows up more than any other provider slug in this log -- 119 of the 496 recorded changes. That is a count of how often the table changed, not evidence that this provider re-prices more often than its listing frequency in LiteLLM's own table.

All 496 changes, newest first

"Evidence" links to the raw price table at the aggregator's own commit before and after the change -- the snapshot that still showed the old number, and the first one to show the new one.