Cheaper LLMs can make your whole workflow more expensive

zeuslac · reddit · 2026-09-21

A widely overlooked cost trap: a cheaper model per call can raise total workflow cost.

Small models do fine on easy cases, but on hard ones their suggestions are weak enough that reviewers read, cross-check, then just do the work themselves — slower than pre-AI. Errors that slip through review cost far more to fix than inference ever cost.
The model bill goes down while work cost goes up, and nobody notices because the model bill is the only tracked number; review time sits elsewhere and later corrections are rarely linked back.

Two follow-on insights: routing by difficulty helps, but its value depends entirely on the price gap between models — a single price cut can remove the reason for the routing layer, so keep it cheap to unwind. And a router can be financially better yet fail on quality: push enough hard cases to the weak model and overall error rate exceeds the manual baseline while the spreadsheet still shows savings.

The fix is unglamorous: one record per task tying model usage, review time and later corrections together, so you see the full cost of completed work rather than the cost of a call.

Original post →

Cheaper LLMs can make your whole workflow more expensive

More from coding & agent