The Cheapest Model I Actually Rely On
Haiku 5.5 at a tenth of a cent per token didn't just save money. It changed which jobs I bother automating at all.
I used to reserve model calls for tasks that felt worth a frontier price. At $0.10 per million input, Haiku 5.5 removed that threshold. Now a throwaway classification, a spam triage, a 'which of these is the duplicate' check all get a model call because the cost is rounding error.
The change in behavior mattered more than the savings. When inference is nearly free, you stop hoarding it, and the small automations you'd have skipped because 'not worth a call' start to exist. That's where the real hours go.
I still don't trust it with the load-bearing reasoning. The effort dial helps, but a $0.10 model making a $10,000 decision is false economy. I use it for the boring 80% and escalate the weird 20% to something bigger.
The lesson isn't 'cheap is best.' It's that price changes the shape of what you build. At this rate, a whole class of tiny agents becomes worth running, and the frontier model gets reserved for the moments that earn it.