The agent worked fine. It cost $4,260.

Understand it one step at a time

The short runs these in order in about 51 seconds. Here they are written out — pick any step to jump the short straight to it.

1. 1

   The agent cost $4,260 last month

   One triage agent. Four thousand tasks. The invoice was six times the estimate.

   billing · ai usage · august

2. 2

   Nobody wrote a bug

   Every task succeeded. There is no incident to find, which is exactly why it ran for a month.

   4,000 tasks · 0 failures · 0 alerts

3. 3

   The API has no memory

   Every step is a fresh request carrying the entire conversation so far. Nothing is stored between calls.

   messages api · stateless · full history each call

4. 4

   Step 12 pays for steps 1 to 11

   The bars are one task. Each one is longer than the last because it carries everything before it.

   step 12 · 24,500 input tokens · one call

5. 5

   Twelve steps. 195,000 input tokens

   Add the bars up. The steps grow in a straight line, so the total grows as a square.

   one task · 195,000 in · 3,600 out · $1.07

6. 6

   Output costs 5x more, and barely matters

   Output is the expensive token and the rare one. The agent barely speaks; it mostly re-reads.

   input $5/M · output $25/M · input is 92% of the bill

7. 7

   Prompt caching reprices the repeat

   A cached prefix reads back at about a tenth of the input rate. Writing it costs 1.25 times.

   cache read 0.1× · cache write 1.25× · 5 min ttl

8. 8

   The same run, 32 cents

   Same agent, same steps, same tokens. $2,976 a month stops leaving the account.

   one task · $0.32 · 4,000 tasks · $1,284

9. 9

   One line puts the bill back

   The cache is a prefix match. Change a single byte near the top and everything after it is new.

   system prompt · rebuilt every request

10. 10

    Nothing errors. Nothing warns

    Caching is switched on, the code is correct, and it has never been read once. Only one number says so.

    cache_read_input_tokens · 0 · every request

11. 11

    Three fixes, in this order

    Freeze what repeats, cache it, and stop dragging dead tool output through every later call.

    freeze · cache · prune

12. 12

    The loop is the bill

    Your agent is not expensive because it thinks. It is expensive because it re-reads.

    ai agent cost · context, not cleverness