An agent hour is not a unit

I have sat in several conversations this year where someone priced an agent hour. An agent working through a night, a certain number of nights per week, at a given amount per hour. It looks like a calculation.

Not one of them could say what the hour contained.

That is not a mark against anyone. It is that the unit does not exist. The hour is a measure borrowed from how consulting work has always been priced, and it does not carry over.

What actually gets billed

A language model is billed per word fragment. Not per character, not per word, but per parts of words, roughly like syllables. You pay for how much information goes in and how much comes out.

Not per hour. And not according to how hard the task feels.

That last part is what surprises people most. A complicated question about a short text is cheap. A trivial question about a large body of material is expensive. The complexity of the task has nothing to do with the price, only the volume.

So an agent hour presupposes an assumption about how much information passes through per hour. That assumption exists in every such calculation, but I have yet to see it written down.

Two consequences that change how you work

Sending in less is cheaper than choosing a cheaper model.

Take an incident. One person tells the model to read the log files from the past twenty-four hours and come back with an analysis. That is an expensive command, because the entire log volume has to be processed.

Another person knows the fault occurred at 18:32, extracts the material around that minute and sends that. Same model, same question, a thousandth of the cost. And a better answer, because the material no longer contains everything else that happened that day.

The difference does not sit in the tool. It sits in someone having understood where the fault was before they asked. The skill is in the input.

And producing text costs roughly five times more than reading it.

That holds throughout, regardless of model tier. It makes translation and content production the expensive steps, since they generate a lot of text. Summary and analysis become cheap, since they take in a lot and leave little.

It is not intuitive. Most people assume the opposite, that the difficult analysis is what costs. But it determines where the money actually lands in a content-heavy project.

One more thing that rarely gets mentioned: Swedish and Finnish draw more word fragments per sentence than English, because compound words break into more parts. The same content therefore costs different amounts per market.

What it does to the model choice

The most common reflex is to pick the most powerful model, because that feels safest. Sometimes it is right. For bounded tasks with known input it almost never is.

Reading a log file needs no reasoning model, it needs to find a pattern in a body of material. There the difference between model tiers can be several-fold in price without being so in outcome.

But note that the model choice is the smaller lever. Halving the price per unit gives you half. Sending in the right material instead of all the material can give you a thousandth.

The question to ask

Not "why does an agent hour cost that much". But what token volume per hour you based it on.

That is not an objection to counting in hours. It is a question about what the hour is built on. And it can be answered by anyone who has actually done the arithmetic.

If the answer comes back as a volume with an assumption about working pace behind it, there is a calculation. If it comes back as a reference to experience from other projects, the hour is an estimate dressed as a unit.

You do not need to be able to redo the arithmetic yourself. You only need to know that it should be possible to show.

See also: An even number over uneven work (series 54) and A check that answers differently is not a check (series 56).