Almost nobody publishes what AI actually costs to run. Vendors quote per-seat pricing, which says nothing about the compute underneath, and pilots get funded on estimates that turn out to be wrong by an order of magnitude in whichever direction is least convenient.
So we measured our own. Every recorded API call against our working repository, parsed out of the local session transcripts and re-priced at published rates: 48,707 calls, and 32 active days of work on the repository itself.
A note on rigour, because it matters more than the numbers. Below, MEASURED means it came out of our own transcripts or git history. MODELLED means we calculated it from assumptions we are stating. We keep these apart because a modelled number presented as a measured one is the failure mode in this entire subject.
Measured: where the money actually goes
Roughly 60% of spend was cache reads — re-reading context the system had already seen. Fresh input, the genuinely new material entering the system, was under 1%. Total for those 32 days: about 8,175 dollars, or around 255 dollars a day.
That is the opposite of the intuition. Most people assume cost scales with how much material you hold. It does not. It scales with how much context each pass re-reads. A company with ten times our documents would not pay ten times as much — but a system that re-reads its whole corpus every night would become ruinous at a few hundred pages.
The design lever is not how much you have. It is how much each pass re-reads.
Measured: the median is trivial, the mean is not
Aggregates hide the real unit. The median session cost 80 cents. The mean was almost fourteen times that, because a small number of substantial working sessions consumed the overwhelming majority of spend — 71 of them averaged around 101 dollars each and accounted for about 88% of the bill.
That tail, not the median, is what an unengineered nightly agent job would look like.
Modelled: what the same work costs built properly
Most of those calls ran on the most capable and most expensive model available, and none were batched. Re-priced on the same token shape — cheaper models for bulk classification, batching for the work that is independent per item, and the expensive model reserved for the small number of genuinely hard judgements — the cost falls by roughly an order of magnitude.
Extended to a nightly workload for one customer, the gap widens. Run as an interactive session left on a schedule, it projects to around 3,000 dollars a month. Designed as a staged pipeline, around 20 dollars.
Be careful with that comparison, and we would rather point at its seams than have you find them. The 3,000 is a measured session cost multiplied by thirty nights — an assumption, not an observation. The 20 is calculated from a pipeline design, and that pipeline has not yet been run in production. The re-pricing also holds the token shape fixed and assumes a cheaper model does the same job to the same standard, which we have asserted from cost arithmetic and not yet demonstrated with a quality evaluation. What the numbers genuinely establish is the size of the prize, not a receipt.
Why any of this matters if you are buying
Because the running cost of an AI system is not a property of the AI. It is a property of how it was built. Two suppliers can deliver the same result, and one of them has a business that survives at your volume while the other quietly does not.
- Ask what the system re-reads on every run, and whether that grows as your data grows
- Ask which model does the bulk of the work and which handles the exceptions — one model for everything is the expensive shape
- Ask whether work that is independent per item gets batched
- Ask for a cost per document, not a cost per seat
- Ask which of their numbers are measured and which are modelled — and watch whether the answer comes easily
One constraint worth knowing, because it invalidates a common assumption: prompt caching does not survive across days. A nightly job always starts cold. Any cost model that quietly assumes yesterday's cache is still warm is wrong, and it is wrong in the direction that makes the quote look better.
The uncomfortable version
If a supplier cannot tell you what their own system costs to run, they have not measured it. That is not a small gap. It means the margin on your contract is unknown to them too, and the usual way that resolves is a renegotiation twelve months in.
We publish ours, with its seams showing, because it is the most useful thing we know that almost nobody says out loud.