# Model choice cheatsheet

Pick the model class from the job, not from a leaderboard.

The model is not the point. The decision is. Work out what kind of decision sits in the middle of your workflow, then buy exactly as much model as that decision needs. Most workflows waste money in one of 2 ways: paying frontier prices for work a small model does fine, or using any model at all for work that is not a decision.

## The table

| Job type | Model class | Why |
| --- | --- | --- |
| Judgment calls and synthesis (triage with tricky edge cases, weighing a lead against your customer profile, deciding what changed enough to mention) | Frontier model | The decision has nuance and the cost of a wrong call is high. You run it once per item, so the price per call barely matters. |
| Extraction and classification at volume (pull the invoice number, tag 5,000 rows, split name from address) | Fast small model | The decision is narrow and repeated thousands of times. A small model gets the same answer at roughly 1/10 the cost and faster. |
| Drafting words a customer will read (reply drafts, follow-ups, report prose) | Frontier model | Tone and judgment show. A stiff or wrong draft costs you more than the API call ever will. |
| Dedup, formatting, math, date logic, sorting | No model. Write code. | These have exactly 1 correct answer. Code gets it right 100% of the time for free. A model gets it right most of the time for money. |

## The 3 questions

1. Is there a judgment call in this step? If no, write code.
2. Does the call repeat at volume with narrow options? Use a fast small model.
3. Will a human customer read the output, or is the call genuinely hard? Pay for frontier.

When in doubt, start with the frontier model to prove the workflow, then downgrade the steps that turn out to be narrow. Downgrading a working system is easy. Debugging a cheap model on a hard job is not.

---

From Agents That Work, lesson 1. advizracademy.com