When you use Claude Code to build a tool that sorts inquiry emails by type, or one that sums up daily reports in three lines, Claude ends up running inside that tool too. This kind of "Claude built into an app" is charged by how much you use it. Even if each item is small, processing hundreds a day adds up in both cost and waiting time.
That's when you face the question of which Claude to use inside the app. Picking the smartest model feels safe, but it's overkill for a job that just sums things up in three lines.
"Claude Haiku 5.5," announced by Anthropic on October 7, 2026, is a new model for exactly this kind of light work. Let's look at what changed, who it helps, and how to ask first.
Contents
What changed: the smallest model got cheaper, faster, and smarter
A model is the part that acts as Claude's brain. Claude's models include Opus for complex work, Sonnet, which is faster and easier to use than Opus, and Haiku, the smallest and lightest. According to Anthropic's announcement, Haiku 5.5 is the cheapest, fastest, and smartest small model it has released so far.
Its cost is about 75% lower on average than the previous Haiku 4.5. In speed, it's the fastest of the models so far when compared at standard settings. Its capability also rose sharply, scoring far above the previous Haiku on a test that evaluates work deliverables.
It's also the first Haiku that lets you adjust, in steps, how deeply it thinks. Haiku 5.5 is a model for handling large volumes of light work, cheaply and quickly.
Who it helps, and how
Haiku 5.5 helps most the people who call Claude from their own apps and tools, in other words, developers who build Claude into apps. If you only talk with Claude in the chat screen, it doesn't directly affect you.
The announcement lists the work it's good at: summarizing, condensing the key points of a conversation that has grown long, searching data, sorting, quick lookups, and customer inquiries that need an immediate reply. On the other hand, when you have Claude handle complex programming, Sonnet 5.5 or Opus 5.5 is recommended.
Even when building an app from zero with Claude Code, you can split things up: leave the building to Sonnet or Opus, and leave the light work that runs every day inside the finished app to Haiku. You can choose the Claude that builds the app and the Claude that runs inside it separately.
How to ask first: describe how heavy the work inside the app is, then have Claude pick
When building an app with Claude Code, it's best to first explain what the Claude inside the app will do and roughly how many times a day it will run, and then discuss which model fits. Once Claude knows how heavy and how much the work is, it can more easily suggest a model to build in.
For an inquiry-sorting tool, for example, the exchange looks like this.
You: I run a small online shop and get about 100 inquiry emails a day. I want to build a tool that sorts them into "shipping," "returns," and "other." Which Claude do you think I should use inside it? Claude: Since it's light work, just sorting into three fixed types, and the volume is high, the cheap and fast Haiku 5.5 seems like a good fit. How about putting only the ones it's unsure about into "other" for a person to check later? You: Let's go with that. Build it with Haiku 5.5 as the Claude inside the tool.
Rather than asking "build it with the best model," consulting from the content of the work, with "100 a day, just sorting into three types. Which one is best?", makes it more likely that a model matched to the cost gets chosen. Choose the Claude to build in together, after describing how heavy and how much the work is.
Summary
With Haiku 5.5, the smallest Claude became about 75% cheaper on average, faster, and smarter than before. Even apps that run large volumes of light work like summarizing and sorting every day become easier to build Claude into without worrying about cost and waiting time.
Instead of leaving everything to the smartest model, choose the model to match how heavy the work is. This way of thinking can be used again and again, every time a new model comes out.
Next time you build an app with Claude inside, explain "how many times, and what, the Claude inside will do," and ask which model is best.