The case for the small, fast model

Anthropic released Claude Haiku 4.5 on 15 October 2025, the small and fast member of its current line. The release is a good moment to make an argument this blog has made in pieces since February: most business tasks do not need the largest model.

Why people over-specify

Because the largest model is the one in the marketing, the one in the benchmark tables, and the one the salesperson demonstrates. Nobody has ever been criticised for choosing the most capable option.

But the difference between a large model and a small one is invisible on the tasks that make up most of a company’s actual usage: summarising a thread, drafting a routine reply, extracting three fields from a document, tidying dictated notes.

What you gain by choosing smaller

Cost, by a factor most people find surprising. Speed, which matters more than it sounds — a tool that answers in one second gets used casually, and a tool that takes fifteen gets used deliberately or not at all.

And, for the self-hosting minority from 4 September, a small model is the one that actually runs on hardware you could own.

Where you still want the large one

Anything with several steps that depend on each other. Anything where the answer must be reasoned rather than retrieved. Anything you will send to a customer without a careful read.

That is the three-tier sorting from April, and it is still the most useful ten minutes a company can spend on this subject.

The practical step

Take your two most frequent AI tasks and run them on the smallest model your provider offers, side by side with what you use now, for a week.

In our experience one of the two is indistinguishable and the other is not. Moving the indistinguishable one down a tier is free money, and the exercise costs an afternoon.