How to Cut Your AI API Costs by 70% Without Changing Models

The first surprising API bill usually triggers the same reaction: switch to a cheaper model. It is the obvious lever and it is rarely the best one, because it trades quality for cost when most bills contain a large amount of pure waste.

Here is what to do first, roughly in order of how much they typically save.

1. Ask for shorter answers

Output tokens cost three to five times what input tokens cost, on every major provider. That ratio is the single most exploitable fact about API pricing, and almost nobody acts on it.

A model asked an open question produces a thorough answer, because thorough is what it was trained to produce. It will explain its reasoning, offer caveats, and summarise at the end. If your application only uses the first sentence, you paid for all of it.

Adding "Respond in under 80 words" or "Return only the classification, no explanation" routinely halves a bill with no change of model and no loss of anything you were using. Check what your application actually consumes from the response — the gap between what you receive and what you use is often most of the cost.

2. Turn on prompt caching

Most providers now cache the repeated prefix of a prompt and charge substantially less for cached tokens — often a tenth of the normal rate.

Almost every production application has a large repeated prefix: a system prompt, a set of examples, a schema, a policy document. If that material sits at the start of every request and does not change, caching it is close to free money.

The requirement is that the stable material comes first and stays byte-identical. Putting a timestamp or a user ID at the top of the prompt breaks the cache for every request, which is a common and expensive mistake. Variable content goes last.

3. Route instead of standardising

Teams pick a good model and send everything to it. Then the bill arrives, and an audit finds that most calls were doing work a cheap model handles perfectly.

The pattern that works is a cheap model deciding what needs an expensive one. Classification, extraction, routing and filtering go to the cheapest tier; only the cases that genuinely need reasoning are escalated. Because the cheap model runs first, it often pays for the entire system by preventing calls that would otherwise have gone to a flagship.

This usually beats simply downgrading, because you keep flagship quality on the work that needed it.

4. Stop sending context the model does not need

Conversation history grows with every turn, and by default all of it goes with every request. By turn twenty you are paying to re-send nineteen turns, most of which have no bearing on the current question.

Summarise older turns rather than resending them verbatim. Drop tool output once its result has been used. If you are retrieving documents, retrieve fewer and better ones — as covered in our RAG guide, more retrieved context frequently makes answers worse as well as more expensive, because it dilutes attention.

5. Batch what is not urgent

Several providers offer a batch tier at around half price for work that can wait, typically returning within 24 hours.

Anything not blocking a user is a candidate: overnight classification, backfills, report generation, embedding a corpus, evaluation runs. The engineering cost is one queue and a callback. The saving is 50% on everything you move.

6. Cache your own answers

Distinct from prompt caching. If the same question is asked repeatedly — and in most products a small number of questions make up a large share of traffic — store the answer and serve it without calling the model at all.

A hash of the normalised input as the key, with a sensible expiry, is often enough. Support and documentation assistants in particular tend to see the same handful of questions over and over.

7. Measure before you optimise

Log token counts per call, grouped by feature. Almost every team that does this finds the distribution is not what they assumed: one endpoint nobody thought about is producing most of the spend, usually because it runs on every page load or retries silently.

Our cost calculator will estimate a monthly bill from tokens per call and volume, and the token counter shows what a given prompt actually costs to send. Both run in your browser.

What not to do

Do not cut quality on the work that matters. The point of reducing cost on routine calls is being able to afford the good model where it counts. A blanket downgrade saves money and costs users.

Do not optimise before measuring. Every list like this one, including this one, is generic. Your bill has a specific shape, and the first thing to do is find out what it is.

Do not confuse cheap with efficient. A cheap model that needs three attempts, or that produces output a human has to fix, is not cheap. Count the whole loop.

For a comparison of what the major providers actually charge, see our guide to the cheapest AI APIs.