What LLM tokens used to cost

The steepest fall we can evidence is mistral-small-latest on Mistral AI: $6.00 to $0.30 per 1M output tokens, a drop of 95% in 445 days.

But the more useful finding is how rarely a published price moves at all. Of the 57 series observed more than once, 41 never changed price between the first observation and the last; 12 got cheaper and 4 got dearer. Tokens have got dramatically cheaper mostly because newer, cheaper models arrived — not because the model you were already paying for was marked down.

Charts are logarithmic: a straight line is a constant rate of change, and a fall from $60 to $0.60 would otherwise flatten everything else into the axis.

How to read this, and what it is not

• These are list prices, standard synchronous tier. Batch, cached-input and off-peak discounts are excluded, so real bills are often lower than the line.
• A date is an observation, not a change date. A snapshot dated March 2024 means the page said that in March 2024 — the price may have taken effect earlier. Rows marked Launch or Price change do carry the vendor's own date.
• Rows marked "LiteLLM, monthly" are not seller-verified. They are read from a dated commit of LiteLLM's aggregated price file, not from the provider's own page — a weaker guarantee than every other row here. Their date is the month a price was first recorded in that monthly snapshot, never the day it changed, and the true change could fall anywhere in the weeks before it. Each one still links to the exact commit it came from, so you can check it yourself.
• Gaps are gaps. Lines connect observations; they are not evidence about the days in between, and a missing model means nobody has evidenced it yet, not that its price held.
• Same name, different thing. Models are quietly revised under a stable name. A flat line can hide a model that got better at the same price — which is itself a price cut per unit of capability, and one this dataset cannot see.
• Prices in CNY are shown in CNY. Converting an old observation with today's exchange rate would invent a number, so we don't.

Take the data

The whole record is one JSON file — every point with its source URL and the quote that locates it on the page. CC BY 4.0, so you may use it anywhere with attribution.

curl https://is.kluuz.de/price-history.json

The LiteLLM-recorded points marked above live in a companion file, kept separate so this one's guarantee never gets diluted: price-history-aggregate.json.

Found a price we got wrong, or have a dated snapshot we're missing? Open an issue — corrections are the point.