There is a default that quietly writes most AI invoices: route everything through the best model you can afford, because the best model is the safest choice. It is the safest choice. It is also, per token, routinely thirty to sixty times the price of a fast small model, and a surprising share of what flows through it is work the small model does identically.
What cheap work looks like
Look inside a month of production traffic and you keep finding the same turns:
- Routing. "Which of five things is this request?" A classification a small model gets right essentially every time.
- Extraction. Pulling a name, a date, an order number out of a message.
- Formatting. JSON in, slightly different JSON out.
- Titles and summaries. Six words over a conversation nobody re-reads.
- Short factual answers already grounded in retrieved text: the retrieval did the hard part; the model is phrasing.
None of these need frontier reasoning. All of them are billed like they do when they ride the default route. The expensive model earns its price on the turns that are genuinely hard (long context, real reasoning, high stakes), which in most products is a minority of the traffic.
How to find the split without guessing
The mistake is doing this by vibes in either direction. The method that works is small and boring:
- Benchmark on your own traffic, blind. Take a few dozen real prompts from your product, run every candidate model over them, and score the answers without knowing which model wrote which. When I did this for the assistant on this site, the fastest model, answering in about half a second, scored within a point of models many times its price on these prompts. That last clause is the whole method. On someone else's prompts, someone else's answer.
- Give every reply a budget that is checked before the call, not summed after the invoice. A ceiling that stops a runaway loop the moment it starts is worth twelve dashboards that explain it a month later.
- Record cost per model call, not per conversation. One user turn can be several billed generations once tools are involved; if you only track turns, the books are quietly wrong and reconciliation against the provider's invoice becomes archaeology.
Why this usually ships invisibly
The satisfying part of this work is what users notice: nothing. The hard turns keep the strong model. The cheap work moves to a model that answers faster, so the product usually gets quicker while the bill drops. The service page here calls that "usually without users noticing", and that phrasing is measured, not modest.
If your AI bill has been drifting up while the product stayed the same, that drift has a shape, and it is almost always this one. Ask the assistant on this site where I would look first: it will tell you, and then it will be billed for having done so, on a small model, with the cost written down.