Haiku, Sonnet or Opus: choose by task, not by size

A customer service team has to draft hundreds of short customer replies a day. Someone proposes always using the most capable model, “so quality is never an issue.” It sounds prudent, but it means paying more and waiting longer on a task that doesn’t need it.

Choosing a model is choosing a trade-off between cost, speed, and quality. The question isn’t which model is best, but which one is enough for this task.

The three families

Claude comes in three model families. Versions change; the logic of each family stays the same:

Family	What sets it apart	What it suits
Haiku	The fastest and lowest-cost	Simple, high-volume tasks: sorting emails, short drafts, pulling data from a form
Sonnet	A balance of speed and capability	Everyday work: writing a newsletter, analysing a spreadsheet, summarising reports
Opus	The most capable, slower and costlier	Multistep complex reasoning: an analysis with many variables, decisions with hard trade-offs

Haiku is not “the one that only fixes typos”: it reasons well for its size. And Opus is not the safe default: it is the right choice when the task really calls for it.

How to choose in practice

1. Look at the task, not the budget or habit. Is it simple and repetitive? Is it routine writing? Does it require reasoning over many interacting pieces?
2. For simple, high-volume work, start with a fast, low-cost model. High-volume tasks are where cost and waiting time weigh the most.
3. Test on a sample. Review twenty or thirty outputs, not one.
4. Move up a model only where quality falls short. If the technical descriptions come out weak and the simple ones are fine, only the technical ones move to a more capable model.

It also works the other way: if a team uses a single model for everything, it pays to split. An operations team with three recurring tasks can assign Haiku to email tagging, Sonnet to the weekly newsletter and Opus to the quarterly staffing analysis. The budget goes where it makes a difference.

Common traps

• Always using the biggest model. More expensive and slower on simple work, with no visible gain.
• Always using the cheapest one. It saves on simple work but fails where reasoning is needed.
• The fast model with no checks at all. Starting with Haiku doesn’t mean skipping the sample review.
• Switching platforms or turning off features to cut costs, instead of choosing the model well.
• Believing a bigger model fixes a bad request. If the prompt doesn’t say who it is for or what it is for, no model can guess (see the prompting domain).

On the exam

Look for the words in the scenario that describe the task: “hundreds,” “every day,” “short replies,” “tight budget” point to a fast, low-cost model; “complex analysis,” “many variables,” “trade-offs” point to a more capable one. The correct option almost always matches each task with its family, or proposes starting with the low-cost model, reviewing a sample and moving up only where needed. Rule out “the most capable for everything,” “the cheapest for everything without review” and options that switch tools instead of deciding.