We publish cost per run to every client every month, which means we have four thousand eight hundred and twelve rows of it to look at. The pattern is duller than the discourse suggests.
Retrieval, not generation
On the median workflow, fetching and shaping context costs more than the completion does. Teams that tune the prompt first are optimising the cheaper half.
The single biggest saving we have made on any account came from caching a lookup that ran four times per ticket and never changed inside a day.
The retry tail
Failed runs cost roughly what successful ones do, and a workflow with a five percent retry rate carries that overhead on every invoice. Fixing the connector timeout was worth more than switching model tier.
Escalations are the cheap outcome
A run that stops early and routes to a person costs a fraction of one that reasons its way to a wrong answer somebody then has to unpick. The human gate pays for itself twice.
Where it landed
Four cents on an average run, across 1,406 runs on an average day. The figure has fallen every quarter since 2024 and almost none of that came from cheaper tokens.
002
/ Keep reading
From the journal