The cheapest model on the price list isn't the cheapest to run We ran the same 25 synthetic invoices through every candidate model: The cheapest listed model answered "rate-limited upstream" on 15 of 25, even when run alone. A model that doesn't answer costs a person's time. The model that read invoices best wrote passport names as one line instead of surname and given names. The grounding check flagged them, and genuine applicants would have waited for a review. So it stays second. On one provider, turning reasoning to "minimal" took about 10 s; turning thinking off took about 2 s. We now pin the measured hosts first. Model choice is an eval, not a price-list decision.