Silent. The agent worked. For a month it answered well and nobody complained about it. Then the invoice arrived, and it was not the number you had been quoted. No error had been raised, no alert had fired, and nothing in the logs pointed anywhere, because from the machine’s point of view nothing had gone wrong. Somebody had written one helpful line at the top of the agent’s standing instructions: today is such a date, the time is such a minute.
What went wrong
A model keeps nothing between messages. Whatever it must know, it is handed again, in full, every single time it is asked anything.
The Cabinet, ArkOne’s reference design for an executive agent, calls that handed-over material the Brief: who the agent is, what it must always do, and where it is working today. Two of those blocks are marked as standing material and expected to arrive unchanged from one request to the next, which is what makes them cheap. Providers charge a much lower rate to re-read text they have already seen, and the Register, the one page recording every setting of the design, carries the ratio: according to Anthropic’s published pricing, a cached read is billed at roughly a tenth of a fresh read.
The condition attached to that discount is the whole of this case note. It covers an opening stretch of the request, from the first word to a marker, and only while that stretch arrives byte for byte identical to last time. The saving is earned by a continuous run of unchanged text, not by the paragraphs inside it.
So a line carrying the time of day, sitting near the top, works as a wall. Everything below it is new material as far as the provider is concerned, on every request: the paragraphs describing the company, the tone, the standing rules, the list of tools, all re-read at full price while the design’s markers sit further down guarding a stretch that never matches. The eleventh paper works through the arithmetic and where the markers belong.
The record
Suppose a company of two hundred people whose agent handles supplier renewals. The company and every figure below are invented for this exercise, and the amounts are modelled from the ratio the Register carries rather than taken from anybody’s invoice. One working session, forty turns, run twice: once with the clock line at the top of the standing instructions, once with the same line moved to the very end.
| Turn | Clock line at the top | Clock line at the end |
|---|---|---|
| First | Everything read fresh; nothing to match yet | Everything read fresh; nothing to match yet |
| Second | Standing blocks read fresh again, because the minute changed | Standing blocks matched, read at the cheap rate |
| Turns three to forty | Fresh every time, thirty-eight more times | Cheap every time, thirty-eight more times |
| Modelled reading bill | USD 4.10, illustrative | USD 0.42, illustrative |
Exhibit 1. Illustrative. An invented company's forty turns, read twice, with the clock line moved from the top to the end.
The two columns describe the same agent doing the same work for the same people. Nobody changed a rule, a specialist or a document between them. The only difference is which end of the standing instructions carries a line that will never be the same twice.
The money is decided at the second turn, in the first minute of the month. From there the agent is either matching its own standing material or it is not, and whichever it starts doing it goes on doing until somebody reads the instructions closely enough to notice a date at the top.
Nothing in the left-hand column looks like a fault while it runs. The answers are the same quality, the agent is no slower, and the only symptom is a number on a document weeks later, by which time the conversation has become one about whether agents are simply expensive.
What happens
The standing part of the Brief is byte-identical to last time, so every re-read is billed at the cached rate.
What your company sees
The action was refused and the reason recorded.
What it means for you
Nobody was overcharged here. The agent asked to have a great deal of text read at full price and the provider read it at full price, so there is nothing to dispute.
What the fault says about the build is worth more than the saving. Somebody added a line at the top to meet a real need, the agent knowing the date, and nothing told them what that position costs. The cheap rate is not a switch to be turned on; a build either keeps qualifying for it on every request or quietly stops, and from the outside the two look identical.
So put the question in terms of position. Ask which parts of what your agent reads are the same on every request, where anything that changes is placed, and what the bill would do if that thing moved. An answer naming the two standing blocks and where the markers sit comes from somebody who has looked.
Whoever built it will tell you the bill is being watched. The useful follow-up is whether two identical runs of one conversation produce the same bill, which turns out to be a harder question than it sounds.
Next: The same tools in a different order, and a bill nobody could reproduce.
Asked plainly
Why would an AI agent suddenly cost ten times more?
Usually because it stopped qualifying for the cheap reading rate. A model remembers nothing between messages, so the whole standing instruction is handed to it again every time, and providers charge a much lower rate for a run of text that is identical to last time. That discount applies from the first word up to the first word that changed. Put one line that changes every time near the top, and nothing after it qualifies.
What is prompt caching in plain terms?
A provider keeps a copy of the opening stretch of a request and charges a fraction of the usual rate to read it again, provided that stretch arrives character for character the same as before. It is a discount on repetition, not a memory. Because it covers an opening stretch rather than a set of pieces, a single changed word early on withdraws the discount from everything below it.
Does a date or timestamp in an AI prompt really matter?
It matters entirely because of where it sits. An agent that needs to know today's date can be told at the end of what it reads, after everything stable, and the cost is one short line at full rate. Told at the beginning, the same line makes every stable paragraph below it new material on every single message, and the cheap rate never applies at all.