LLMPrice indices · 2026-08

Which of these is your workload

The four shapes are not categories of business, they are ratios of text in to text out. Pick the one that sounds like what you run and price it on today's rates.

• Answering questions from your documents

  Long context read repeatedly, short answers out. The Retrieval shape, and the one that benefits most from caching.

• A customer support chatbot

  Roughly as much text out as in, back and forth. The Chat shape.

• Writing content, reports or summaries

  Short brief in, a great deal out. The Content shape, and the most exposed to output rates.

• An agent working through a task

  Many steps over a stable prefix, little written back each time. The Agent shape.

Downloads and citation

Full series, JSON  ·  Full series, CSV

These are the living files: they lengthen by one reading each month. To cite a figure, use the dated issue, whose own downloads are frozen with it.

Showing a reading on your own page rather than quoting one: the badge carries any of the four values and updates when the monthly reading does.

How to read these indices

Each index prices a fixed basket, the LLMPrice 12: two models per provider, the newest member of the provider's own flagship line and of its value line. An index moves only when one of those twelve is repriced. It does not track every model on the market, and it does not track the wider repricing record on the site's ticker, which includes models outside the basket.

A new model entering the basket cannot move the index. Substitutions are chain-linked, so only a price change on a model present in consecutive readings registers. This is deliberate: a cheaper replacement reaches your bill only if you migrate to it, so it must not appear here as though prices fell on their own.

Providers rarely reprice a model in service. Only 10 of the 25 monthly links in this series carry any change at all, and the long flat stretches in the chart are the finding, not a gap in it: the price of what a team already runs barely moves, while the market's headline prices fall by replacement.

That gap is what these indices exist to measure. A busy repricing ticker beside a flat index is the market working exactly as observed: individual models change price occasionally, new models arrive cheaper constantly, and neither reaches an existing bill without a decision. The basket, the line assignments and their sources are on the methodology page.

Deflation by cacheable share of input

A standing section, reported every issue. The four profiles differ in how much of their input can be served from cache. Ordering the profiles by that share shows whether price movement has tracked it.

Movement is ordered by cacheable share of input this issue: the more of a workload's input can be cached, the further its index has fallen. This is a correlation across four profiles, not a demonstration of intent. It is reported each issue whether or not the ordering holds.

History

The series began 2024-07 and lengthens by one reading each month. 26 monthly readings so far, chat profile shown. It is short, and will stay short for a while.

The LLMPrice 12, as priced this issue

Substitutions

Every basket change since the series began, with the linking factor applied so the change does not appear as a price movement.