EP5 Choosing the right model: dozens of times cheaper, no worse work

In EP4 our AI team got a production line. Once you use it every day, the question that always follows is cost. The good news: most of it can be cut sharply without worse work, simply by not using the most capable model for everything.

The problem: hiring an expert to do the photocopying

If a senior engineer did every job in a company, the work would be good and far more expensive than it needs to be. Searching folders or summarising reports does not need that level. Using the largest model for every task is the same mistake.

Principle: use the capable model for work that needs thinking, and a small one for work that reads a lot but decides little.

The three tiers

Anthropic API prices as of October 2026 (US dollars per million tokens):

Model	Input	Output	Good for
Claude Opus 5.5	$4	$20	planning, system design, hard bugs, important reviews
Claude Sonnet 5.5	$2	$10	everyday coding, general agent work
Claude Haiku 5.5	$0.10	$0.50	searching files, summarising logs, classifying, high-volume repeat work

The Haiku price applies to prompts up to 100K tokens. Above Opus there is also Claude Fable 5.1 ($10 / $50) for the hardest and longest work.

Notice two things. First, output costs 5 times input on every model. Second, Haiku is 40 times cheaper than Opus.

A worked example: sending a helper to search code

Say a subagent (EP3) has to find everywhere a function is called. It reads 80,000 tokens of code and writes back a 2,000-token summary.

Model	Input	Output	Total
Opus 5.5	0.08 × $4 = $0.32	0.002 × $20 = $0.04	$0.36
Sonnet 5.5	0.08 × $2 = $0.16	0.002 × $10 = $0.02	$0.18
Haiku 5.5	0.08 × $0.10 = $0.008	0.002 × $0.50 = $0.001	$0.009

A search like this needs no deep thinking, and Haiku does it well enough. Fifty times a day, that is about $18 against less than $0.50.

Setting it up in Claude Code

The main agent switches model with the /model command. Use the capable one for planning or hard problems, and step down when the work is routine.

Each subagent can name its model in its agent file:

---
name: finder
description: Find where code lives and where it is called from. Use when a search spans many files.
tools: Read, Grep, Glob
model: haiku
---

The main agent thinks and decides on Opus, and searches go to finder on Haiku, giving the saving from the example above automatically. A security plugin in Claude Code’s official plugin marketplace does the same: its code explorer is set to Sonnet rather than the largest model.

Turn down effort before switching model

Recent models have an effort dial (low, medium, high, xhigh, max) for how deeply to think before answering. Low effort means shorter thinking, fewer tokens and faster answers.

The recommended order: if a simple job feels expensive, lower the effort first, and only then consider a smaller model. A capable model at low effort often still beats a small model at high effort, and you change one thing at a time.

Hidden costs people miss

• Answers longer than needed. Output costs five times input. Preambles, repeated summaries and decorative tables are all money. The machine this article was written on has a personal rule in its CLAUDE.md (EP1) to answer briefly, dropping pleasantries and preambles, but never cutting what keeps the work correct.
• A handbook that keeps changing. What is sent every time, like CLAUDE.md, is cached, and cached reads cost roughly a tenth of normal. Edit the file constantly and the cache has to start over.
• The cheap model that needs redoing. If a small model gets it wrong, needs three rounds of fixes and extra human checking, it ends up costing more than one pass on the capable one. Measure cost per finished job, not per call.

Try it in 10 minutes

1. Create the finder subagent above with model: haiku
2. Ask the same code-search question twice: once letting the main agent search, once sending it to finder
3. Compare the answers. If they are equally complete, send searches to finder from now on

Where a person is needed

A person decides which work must not fail, and that work goes to the capable model. A person also measures whether moving to a smaller model really gives the same result, rather than guessing from the price list. Savings that break the work and need fixing later are not savings.

Next episode

So far our AI works with files and commands on the machine. EP6 gives it hands: access to other systems such as databases, websites or company tools, through one standard called MCP, the first door to working with other systems and other AIs.