Guides

Reading a token bill without overstating it by 10×

Input, output, cache reads and cache writes bill at very different rates. Summing them into one number and pricing it at the input rate is the most common way to be badly wrong about what a session cost.

4 min read

A session reports four token counts, and they bill at four different rates. The single most common mistake is adding them up and multiplying by the input price.

The four numbers

  • Input — fresh context, charged at the full input rate.
  • Cache write — context stored for reuse. Charged above the input rate, because it costs more to write than to send.
  • Cache read — context served from that cache. Charged well below the input rate, often by an order of magnitude.
  • Output — what the model generated. The most expensive per token, and usually the smallest count.

Why summing them goes so wrong

Cache reads dominate the total in a long session — they are frequently 99% of it. Pricing that at the input rate overstates the session by close to an order of magnitude. Doing the same thing in the other direction, pricing everything at the cache-read rate, understates it just as badly.

There is no single number to report here. There are four, and a price list that changes per model and per provider — which is why Inlinr shows counts and not money. A wrong figure about cost is worse than no figure: one gets checked, the other gets quoted.

If you want a rough figure anyway

Multiply each of the four by its own current rate for the model you used, and add those. That is the whole method. Anything shorter is a guess with a decimal point on it.

Inlinr stores input, output, cache reads and cache writes separately and never sums them into one figure, so you can apply whichever price list is actually yours.

More guides