Most AI bills are not big because the work is hard. They are big because of lazy defaults: the most expensive model on every task, the biggest context every time, and no thought about which jobs actually need the premium option. Here is how to spend far less without your results getting worse.
Stop using the frontier model for everything
This is the whole game. Match the model to the task. The hardest coding and highest-stakes writing deserve the expensive model. Summaries, classification, formatting, drafts, and bulk work do not. A cheaper model, or a good open one, handles the easy 90 percent for a tiny fraction of the price, and nobody can tell the difference in the output.
Paying frontier prices for easy tasks is like taking a helicopter to the corner shop. It works, and you are an idiot for doing it.
Trim the context
You pay for every token you send. Stuffing the entire document when the model needs two paragraphs is money set on fire. Send the relevant slice, not the whole file. This alone can cut costs sharply on retrieval-heavy work, and it often improves answers, since the model is not distracted by noise.
Cache and reuse
- Cache repeated calls. If you are asking the same thing many times, store the answer instead of paying for it again.
- Reuse a fixed context. Many providers charge less for a prompt prefix they have seen before. Structure your prompts so the stable part comes first.
- Batch where you can. Non-urgent work often qualifies for cheaper batch pricing.
Measure before you optimize
You cannot cut what you do not track. Log which models and how many tokens each part of your workload uses, and you will almost always find one or two jobs quietly eating most of the bill. Fix those first. The rest is rounding error.
The takeaway
Route hard tasks to expensive models, easy tasks to cheap ones, send less context, cache the repeats, and measure. Do those five things and halving your bill is not optimistic. It is what happens when you stop paying for a feeling of safety on work that never needed it.