2026-07-08 — 6 MIN — WORKFLOWS
Which AI model should you hire? The honest routing table
Treat models like employees with different salaries. The most expensive hire is wasted on routine work, and the cheap hire fails expensive-to-debug tasks. The skill is routing: matching the job to the cheapest model that reliably completes it.
Below is the routing table we run at Robots on Payroll, for the newsletter itself and for vibe-coded products. Prices move; the logic doesn't.
| Job | Hire | Why |
|---|---|---|
| Quick questions, edits, summaries | Cheapest tier (Haiku-class) | Same accuracy on simple tasks; 10-50x cheaper |
| Everyday coding, drafts, analysis | Mid tier (Sonnet-class) | The 80% workhorse; intro pricing often makes this the value play |
| Hard debugging, architecture, long autonomous builds | Frontier tier (Opus/Fable-class) | One-shots what cheaper models retry; cheaper per OUTCOME on hard work |
| High-volume pipelines (classification, extraction) | Cheapest tier + batch APIs | Batch discounts commonly halve the bill |
| Planning-then-building | Frontier plans, mid tier executes | The orchestrator pattern: one worked example cut costs 74% |
The counterintuitive rule: premium models are for when being right the first time is worth more than being cheap per token. One audit we covered found a frontier model cost $4.40 per real bug found vs $17.55 with the mid tier. Double the token price, a quarter the price per outcome. Route by outcome, not by sticker.
The equally counterintuitive inverse: on simple tasks, rigorous testing keeps finding that premium models are no more accurate, just slower and pricier. Paying frontier rates for routine work is the most common money leak we see.
GET THE WEEKLY VERSION
Twice a week: real workflows, routing math, and receipts. Free.