The shape of it
One month cost half of everything
Here is the whole history. Apr is not a trend, it is an accident —
and an instructive one, because nobody noticed it while it was happening.
Mar$26.18
Apr$309.89
May$75.86
Jun$61.13
Jul$86.67
Aug$78.88
Sep*$10.80
* Sep is a partial month (28 days). Bars are scaled to Apr. Every figure is
the sum of the usd field across the usage ledger for that month.
What a blowout looks like from the inside
Apr was $309.89 across 23,065 calls — 48% of six months' spend in
four weeks. Two things caused it. The system was routed to a reasoning model priced at
$30 per million in, $180 per million out, and it was running multi-agent
strategy panels that fan a single question out to several expensive models at once. 1,040
calls to that one model cost $93.51 by themselves.
Nothing was broken. No alert fired, because no cap was breached — the global daily
ceiling was never close to binding. The bill was simply the honest price of a routing
decision nobody had costed, made once and left in place for a month.
Same model class, a generation later, is a sixth of that price. Most of what
looks like optimisation above is just the market moving underneath us — worth saying
out loud, because it is not a skill we can claim.
The actual finding
The waste was not the model price. It was a panel that kept failing.
Every cost conversation about agents is about picking a cheaper model.
Ours turned out to be about something duller: a strategy panel. Each strategy question went to
three models at once and their answers were merged. The seats failed thousands of times, often
all three together, and the ledger priced every failed call.
Failed calls — 14,069 of 31,934, at $108.11
| Panel model | Failed calls | Priced at | What we know |
|---|
| A preview model | 4,899 | $23.85 | never answered: not one live row in the ledger, and it kept its seat into July |
| A general workhorse | 4,546 | $27.46 | not dead: it answered live throughout. Nearly all of these fell in April, when all three seats failed together |
| A frontier model on an empty account | 4,453 | $54.16 | the account's API credits ran out that spring and the seat stayed on the panel |
| Three panel models | 13,898 | $105.47 | 98.8% of all failures |
The preview model is the one that stings. It sat on the panel from March to July
and failed 4,899 times without returning a single answer of its own. It did not look
broken from the outside: the panel still handed back something every time — the other seats’
answers when they worked, canned text when they did not.
Why a failing panel is invisible
A panel is designed to absorb a missing seat. One seat fails, the others answer, and the
merge still reads like an answer. When every seat failed at once, the agent fell back to
canned text — and canned text reads like an answer too. Nothing upstream could tell
“the panel answered” from “one seat has not answered since March.” The
system reported healthy, and every run still added priced rows to the ledger.
We only found it by parsing the usage ledger by status rather than by model — which is
not a view any dashboard offers by default, because the interesting number is a ratio between
two columns nobody thinks to divide.
Read this before quoting the number
The daily rate is not pure efficiency
Sep's $10.80 over 28 days works out to about 39 cents a day, and
it would be dishonest to present that as pure efficiency. Much of the agent's reasoning moved
onto a subscription-authenticated path that this ledger does not bill at all. Some
capability got cheaper. Some capability simply stopped appearing on this particular invoice.
The ledger has other blind spots we know about: a dozen scripts that call models directly
instead of through the router, all gateway traffic, and any model whose price the system
doesn't recognise, which logs as $0.00 rather than as unknown. A cost figure with unlisted
exclusions is a marketing number. These are ours.
What transfers
Four rules, if you are running agents on a budget
Rule 01
Audit by status first
Cost per successful result is the only figure that means anything. Ours was 20%
higher than the sticker price on every model we used, and no per-model view would ever have
shown that.
Rule 02
A seat that never answers is not a second opinion
It is a line item. Pull a panel seat the day it stops answering, make an empty panel fail
loudly instead of returning canned text, and report which seats answered, not just that
something did.
Rule 03
Price the routing, not the model
The Apr blowout was one configuration change nobody costed. Caps didn't catch it because
caps guard against runaway volume, and this was ordinary volume at six times the unit
price.
Rule 04
Cut waste, never capability
Every saving above came from deleting calls that produced nothing, or from prices falling.
Not one came from giving the agent a weaker model for the same job. A cheaper agent that
makes worse decisions is not a saving, it is a slower way to lose money.
The offer
What the $649.41 bought
Six months of an agent running a business end to end — and, since 28 Aug,
2,388 cold emails for a paying-client campaign that produced its first held meeting. That story
is published with the same discipline as this one: full denominator, failures included.
Rick Acquisition — $800/month
Figures re-summed nightly from the agent's usage ledger — 31,934
rows as of 28 September 2026, every one parsed, nothing sampled. Model names are generalised because the specific
versions date faster than the lesson does. Prices quoted are those in effect when the calls were
billed and are not current vendor pricing. No model vendor is affiliated with or has endorsed
this page.