Haiku, Sonnet or Opus: route AI work by how much judgment it needs

Most of the AI bills I look at are high for one reason… every job goes to the biggest model. The intake form, the email sorting, the weekly summary, all of it runs on the most expensive tier because that is what was selected on day one and nobody went back.

Anthropic ships Claude in three sizes. Haiku is the small, fast one. Sonnet is the middle. Opus is the large model for the hardest reasoning. The names and version numbers change every few months, BUT the shape stays the same, and the shape is what you should plan around.

Think in jobs, not models

The question is not “which model is best.” The question is how much judgment a specific job needs, and how many times a day it runs.

• Sorting and tagging - Does this email belong to billing, scheduling or sales? Is this form a real lead or spam? Small model. It runs hundreds of times, the answer is one word, and a wrong answer is cheap to catch.
• Pulling fields out of a document - Name, date, invoice total, policy number. Small or middle, depending on how messy the documents are.
• Writing and summarizing - A draft reply, a meeting summary, a first pass at a proposal section. Middle model. This is where most small-business work lands.
• Weighing things against each other - Reading a contract against your intake notes and your own terms, or untangling a scheduling conflict across three calendars. Large model, and usually not very often.

If a job is high volume and low judgment, it goes small. If it is low volume and high judgment, it can afford to go big. The expensive mistake is the high volume job running on the big model, because that cost repeats all day.

What a routed workflow looks like

Here is a shape I use a lot for an owner’s inbox.

• The small model reads every inbound email and labels it: lead, client, vendor, billing, noise.
• Noise gets archived. Vendor and billing get filed.
• Leads and client emails go to the middle model, which drafts a reply using the business’s own FAQ and pricing notes.
• If the middle model flags something it is unsure about (a complaint, a legal question, a refund request), that one item goes up to the large model for a more careful draft.
• A person reads every draft before it goes anywhere. Nothing sends itself.

The large model touches maybe a handful of emails a week in that setup. The small model touches all of them. That is the right way around.

I wrote more about why the review step stays in, even when the drafts are good, in The owner’s inbox: an AI triage setup that never sends without you.

Escalate on evidence, not on worry

The most common objection I hear is “what if the small model gets it wrong?” Fair. It will, sometimes. So does the large one.

What I do instead of guessing:

• Pull 20 to 30 real examples of the job from the last month, including a few weird ones.
• Run them through the small model and write down what it got wrong.
• Only if the misses matter, try the middle model on the same set. Same for the large one.

That takes an afternoon and it settles the question with your own data. Most of the time the small model handles sorting just fine, and the middle model handles drafting just fine. The big model earns its place on the few jobs where it clearly does better.

Keep that test set. When a new model version comes out, you rerun the same examples and you know in an hour whether switching is worth it.

Two easy savings most people skip

Batch anything that is not urgent. Anthropic’s Message Batches API processes requests asynchronously at 50 percent of the standard price, and most batches finish in under an hour (Anthropic docs). End of day summaries, bulk tagging of old documents, a weekly cleanup of the CRM… none of that needs an answer in two seconds.

Cache the parts that repeat. If every request starts with the same long set of instructions, your price list, or your policy manual, prompt caching lets the model reuse that prefix instead of paying full price to read it every time. I walk through how that works in Prompt caching for owners.

What this looks like on a bill

I’m not going to give you a made-up percentage, because it depends entirely on your volume and your mix of jobs. What I can tell you is the pattern. When a workflow moves its sorting step from the large model to the small one, that step’s cost drops by a lot, and the overall bill follows, because sorting is usually the highest volume step in the whole chain.

If you want the bigger picture on what an automation costs to build and run, What AI automation actually costs a small business has the numbers we quote.

The short version

• Match the model to the job, not to the company.
• Small for sorting, middle for writing, large for weighing.
• Test with your own examples before you move anything up a tier.
• Batch what can wait and cache what repeats.

If you’d like a second set of eyes on how your current AI tools are set up, send me a note and I’ll take a look. Please let me know if you have any questions, thanks!