Skip to Content

Which Claude model and effort level should you pick per task?

The model decides how much Claude knows, the effort level how hard it works. A rule per task saves money.
November 18, 2026

Claude has four models (Haiku, Sonnet, Opus and Fable) and five effort levels, from low to max. If you run everything on the same model, you overpay for simple work or get too little quality on hard work. A fixed rule per task keeps costs under control without giving up on the result.

We follow a fixed circle of makers and communities who work with AI every day themselves. This rule came up in a post by Ben van Sprundel in the Ben AI community. Note: the post also sells the author's own add-ons, and the figures in it come from the author himself.

Model and effort are two different dials

The model decides how much Claude knows. The effort level decides how hard it tries: how many steps it takes and how much it decides on its own. One does not replace the other. Sonnet on max still does not know what Opus knows, and Opus on low skips steps it could have taken.

Which model for which task

  • Haiku for simple, repetitive work that is easy to check: summarising, sorting, routing. Not for work where Claude has to use its own judgement.
  • Sonnet when you can say in one sentence when the task is done, such as putting a sales call into your standard proposal template.
  • Opus on medium when you are not sure. According to the post, Anthropic itself calls that the default answer for most tasks.
  • Fable for big, open-ended jobs where a mistake really costs money. According to the post, Fable costs twice as much as Opus per token.

Which effort level

Go through four questions and stop at the first yes:

  1. Will you go back and forth, or is it a simple task? Then low.
  2. Would a hidden mistake cost money, a client or a wrong decision? Then high.
  3. Are you handing off a long, hard and critical job to walk away from? Then extra high.
  4. Anything else: medium, the default.

The author advises against max: a lot of extra cost for little gain.

Result falls short? Fix it in this order

First ask whether Claude did not know enough or did not try hard enough. Then work through this order:

  • Improve the prompt first. More effort does not fix a bad brief.
  • Let Claude check its own work against an example or a list of requirements.
  • Raise the effort by one level.
  • Only then move to a bigger model, and at a lower effort level.

Our view: choosing a model is a process decision, not a habit. Record for each recurring task which model and level belong to it, so nobody has to think about it every day.

What you do now

  • Pick three tasks your team does with Claude every week and note a model and effort level next to each.
  • Run one task on a lighter model for a week and compare the result with what you are used to.
  • Switch models in a new chat, with a short handoff of what has already been done.

How it works under the hood

Subagents, the helper agents your main chat sends off to read and summarise, run on the same model as the main chat by default. According to the author they can account for up to 90 percent of usage on some tasks; so put reading tasks on Sonnet or Haiku. A skill can get its own model setting, which according to the post only works in Claude Code. The cache, the saved copy of your chat that lets Claude re-read it cheaply, is tied to one model. Switch models in the middle of a chat and the first message pays for the full context again.

This news touches chapters 9, 19 and 24 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.

What you can do with this

Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.

Written in response to a post by Ben van Sprundel in the Ben AI community (14 November 2026, members only).