Main idea

Two flows run through a skill and they don't carry the same rate. What you give the model to read — files, databases, instructions — counts at face value. What it produces — written analyses, records, summaries — counts about five times more.

A skill that reads 100,000 tokens and generates 20,000 therefore doesn't spend 120,000 units but 200,000: half its consumption comes from the 20,000 it wrote. The audit of competitor_analyze shows the same reversal at real scale — 206,000 tokens read against 38,000 generated, and yet generation weighs nearly half the 478,000 units of the total.

Raw token volume and cost therefore don't rank the steps in the same order.

Why it matters

This changes the order of suspicion when you look for where a quota went: you look first at what the agent writes, before what you gave it to read. A short generation step can cost more than a massive read.

It also makes operations of different natures comparable — opening files, writing a report, calling a script — in a single unit.

Nuances and limits

The factor of five is an average rate, exact for certain models at a given date. Pricing grids move, and a particular contract may have another one.

And the weighting says nothing about latency: the most expensive step isn't necessarily the one that makes you wait.

Open questions

• How do you weight an input of which part is cached and billed at a lower rate than the rest?