September 21, 2026
Cheap model makes work more expensive: reviewer ends up redoing complex cases themselves
A smaller model is cheaper per call, but makes the full task more expensive. On difficult cases, its suggestion is weak: a person rereads it, cross-checks it against documents, and does the work themselves, while errors that slip past review later cost more than the model's entire bill.
A smaller model is cheaper per call, but makes the full task more expensive. On difficult cases, its suggestion is weak: a person rereads it, cross-checks it against documents, and does the work themselves, while errors that slip past review later cost more than the model's entire bill.
As long as review time and late-stage fixes are tracked in other systems, token savings show up only on the model bill.
