Notes on working with AI
The small model is the workhorse
Most work in an AI workflow is small, routine calls. Paying top-model prices for them is the main waste. Use a big model to plan and small models to do the volume.
The release worth watching is the cheap one
Model announcements focus on the smartest model. For anyone running AI inside a business, the more consequential one this month is the small one: Anthropic's Claude Haiku 5.5, which the company describes as its fastest and cheapest model, for high-volume, cost-sensitive work.
The reason is how work is actually distributed. Open any working AI setup and count the calls. Most are small and routine: sort this message, pull these fields from an invoice, summarize this thread, decide which team a request belongs to. A minority involve real judgment. Yet most setups send everything to the same strong model, because it's the one that was configured first.
What small means in practice
A smaller model has fewer parameters, so each token takes less compute to produce. The consequences are speed and price. Anthropic lists API pricing for Haiku 5.5 from $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100k tokens, against $1 and $5 for Haiku 4.5. Anthropic says that works out around 75% cheaper on average. It is also the first Haiku with an adjustable effort setting, so you can trade cost against depth per task.
Anthropic's own examples fit the pattern: summaries, classification, database queries, live customer support, and subagent work alongside larger models. Prices and claims like these move, so check the pricing page before you budget.
The architecture: one planner, many workers
The design that follows is simple. A strong model decides what needs doing and breaks it into pieces. Small models carry out the pieces, in parallel, at a fraction of the cost. The strong model checks the result.
This is already how subagents work in Claude Code. A subagent can be pinned to a smaller model with one line in its definition:
model: haiku
You can also switch your whole session with /model haiku, after running claude update, since the new model needs a recent version.
Where it breaks
Small models struggle when a task needs several steps of judgment, when instructions are ambiguous, or when an early mistake compounds into later ones. Anthropic itself says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding. A small model is fast partly because it does less reasoning per token, and some tasks need that reasoning.
So the rule is not to use the cheapest model everywhere. It is to stop using the strongest model by default. Start every recurring task at the small end and move it up only when output quality forces you to.
What to do tomorrow
Write down the tasks you run every week and paste this into any Claude chat:
I will list tasks that recur in my work every week. For each one, say whether a small, fast model is enough or whether it needs a strong model, and give a one-sentence reason. Sorting, data extraction and summarizing usually fit a small model. Multi-step judgment or ambiguous instructions usually need a strong one. My tasks: ...
Take the first group, test it on ten real examples, and compare the output to what you get today.