Which AI model should your team use?

Our plain-English guide to matching AI models to jobs. Where to start, when to move up or down, and how to stop paying top-tier prices to sort an inbox.

Most teams we work with pick one AI model and use it for everything. Some pay top-tier prices to sort an inbox. Others hand a small, fast model a job it was never going to get right and decide AI doesn't work for their business.

Every major AI company sells its models in tiers now. Claude has four, called Haiku, Sonnet, Opus, and Fable. ChatGPT, Gemini, and Copilot sell the same kind of ladder under their own names. On a short request, Claude's top model costs 80 times what its smallest one does.

What the same request costs on each model

  • Haiku1x$0.10 in, $0.50 out
  • Sonnet20x$2 in, $10 out
  • Opus40x$4 in, $20 out
  • Fable80x$8 in, $40 out
The same request on each Claude model, compared with Haiku. Dollar amounts are Anthropic's list prices per million tokens, the unit AI companies use to measure text going in and coming out, as of November 2026.

We use Claude as the example because it's what we build on most. The same approach works in ChatGPT or Gemini. In most chat apps you pick the model from a menu at the top of the chat window, and the thinking setting sits next to it.

Start one step below the top

Start every new kind of task on the second-best model your tool offers, with the settings left alone. For Claude, that's Opus. Then look at what comes back.

If the work is well defined, with clear instructions and a set format, move it down a tier to Sonnet. If it's sorting emails, forwarding requests to the right person, or copying details like totals and dates out of documents, move it down again to Haiku. If the answers are wrong or shallow even with good instructions, move up to Fable.

Teams usually start on whatever was cheapest or whatever came with their plan, get a bad result, and stop there. Starting near the top tells you quickly whether a task is easy or hard.

What each model is good at

Haiku handles thousands of small, similar items, where each answer is short and a wrong one is cheap to spot. It's good at tagging inbound emails as new jobs, warranty calls or spam, or pulling the invoice number and total out of a stack of supplier PDFs. If one in a hundred comes out wrong, someone catches it in a minute.

Use Sonnet for well-defined work and anything a person is waiting on, like a weekly sales report in a format you already use, a reply to a hotel guest asking about late checkout, or a first draft of a slide deck. It's fast, and it's half the price of Opus.

Opus is where we put work when a mistake is expensive or hard to see. That could be a financial model built from five years of sales, or a brand plan pulled together from sales data, retailer feedback and research. Anything new to your team should start here too, until you know how hard it is.

Save Fable for the hardest jobs. Use it for long jobs the model works through on its own, or when Opus on its highest setting still gets it wrong. It costs twice what Opus does.

Turn up the thinking before you switch models

Most AI tools now have a setting for how long the model works on a problem before it answers. Claude calls it effort and offers low, medium, high, and settings above that. Other tools call it thinking or reasoning. Most teams never change it, and it's usually a cheaper fix than switching models.

Anthropic's published numbers for Opus show what each setting buys. Going from high down to medium meant about two or three more wrong answers per hundred, and cost 30% less. Going one step above high meant one or two fewer wrong answers per hundred, and cost two and a half times as much.

What each thinking setting costs on Opus

  • Medium0.7x2 or 3 more wrong answers per hundred
  • High1xThe comparison point
  • One above high2.5x1 or 2 fewer wrong answers per hundred
Cost and accuracy on Opus at each thinking setting, compared with high. From Anthropic's published test results.

When a model falls short, we work through three steps in order.

  1. Give it a way to check its own work, like a finished example, a checklist, or the right answers for a few sample cases.
  2. Turn the thinking up one level.
  3. Move up a model. Start the bigger one at a lower thinking setting than you were using, because every minute of thinking costs more on a bigger model.

If the model doesn't understand commercial leases, it won't figure them out by thinking longer. Move up a model, or give it better material to work from.

Judge the cost of the finished job

AI companies price by the amount of text going in and out. The number to watch is what one correct, finished piece of work costs, including the time someone spends fixing it.

A cheap model that needs three attempts and an hour of cleanup costs more than a bigger one that gets it right once. And paying Fable prices to label 5,000 emails is money thrown away.

In one of Anthropic's tests, two problems out of twenty used 43% of the budget. So when you're working out what an AI workflow will cost, test it on your worst real examples. For a property team, that's the lease with six amendments.

Start a new chat when the work changes

Every time you send a message, the model rereads the whole conversation. On a pay-as-you-go plan, the spreadsheet you pasted in on Monday gets paid for again with every reply on Friday. On a monthly plan, it eats into your usage limit.

When the task changes, start a new chat with a short summary of where you got to. It's also the right moment to switch models.

Use more than one model on bigger jobs

The systems we build almost never run on one model, and neither does our own work. A strong model makes the decisions, and cheaper models do the reading and sorting underneath it.

Take a property manager getting 400 maintenance emails a week. Haiku reads every email and pulls out the unit, the issue, and how urgent it is. Sonnet drafts the reply to the tenant. A person approves anything that commits money. Once a month, Opus reads the summaries and writes the owners' report, with the repeat problems and what they cost.

Where we'd start each job

The jobStart withWhy
Tagging and forwarding at volume, like maintenance requests by unit and urgencyHaikuShort answers, and a wrong one is easy to spot
Copying details out of documents, like totals from supplier invoicesHaikuSimple, repetitive, and easy to check
Replies someone is waiting on, like a hotel guest asking about late checkoutSonnet, low or medium thinkingSpeed matters most
Everyday documents, slides, and spreadsheets in a format you already useSonnetThe scope is clear
A complex financial model, like five years of sales broken out by productOpusMistakes are expensive and hard to spot
A plan pulled together from a large pile of sourcesOpusIt has to weigh a lot of material
Your hardest project, or a long job that runs on its ownFableThe most capable model, at twice the price of Opus
Anything newOpus, default thinkingThen move up or down from what you see

Before any of this, pick one tool

Picking a model comes second. The companies we see getting the most from AI agreed on one approved tool early, wrote down what can and can't go into it, and trained people on the work they do every day.

Then test it on your own work. Pick five real tasks from last week, including the hardest one. Write down what a good result looks like. Run each on two models at two thinking levels, and compare the results against the cost. It takes an afternoon. Rerun it when a new model comes out, because prices and names keep changing.

If you want help setting this up, book a free audit and we'll look at what your team runs today.

Prices and test results are from Anthropic's November 2026 guide to choosing a Claude model.

More guides.

Find out what we'd build for you.

The free audit is a 30-minute call. We ask how your team works today and tell you where we'd start.

Book a free audit ↗︎