How to Reduce LLM API Costs
AI pricing guide · updated 2026-08-07
Most advice on reducing AI costs starts with trimming prompts, which is close to the least effective thing you can do. The levers differ by orders of magnitude in impact, so it is worth working through them in the right order.
1. Use a cheaper model for the work that allows it
The models we track span roughly 1,600× in blended cost, from $0.02 to $33.00 per million tokens. No other lever comes close to that range, which makes model selection the entire ballgame.
The practical pattern is routing rather than switching. Classification, extraction, formatting, routing and simple summarisation are frequently indistinguishable between a frontier model and one twenty times cheaper. Send those to the cheap model and reserve the expensive one for work that genuinely needs it.
Test this rather than assuming it. Run a sample of real traffic through a cheaper candidate and compare outputs, the answer is often that most of your volume never needed the premium model.
2. Cut output, not input
Output costs about twice as much per token as input on the median model, and 3× or more on 117 of the 271 we track. So a token saved on the output side is worth two or more saved on input.
Concretely: set a maximum output length, ask for structured or terse formats rather than prose, stop asking models to restate the question before answering, and avoid reasoning modes on tasks that do not need them, since hidden reasoning tokens bill at the output rate.
3. Turn on caching and batching
If your requests share a long stable prefix, prompt caching cuts that portion to roughly a tenth of its cost. Check the break-even first, cache writes cost about 25% extra, so it needs reuse to pay.
For anything not user-facing (nightly enrichment, backfills, evaluation runs) batch processing typically halves both rates in exchange for asynchronous delivery. There is no quality trade-off, which makes it unusually easy to justify.
4. Send less context, but send it deliberately
This is the lever people reach for first and it is genuinely worth doing, just fourth. Retrieve the few passages that matter instead of stuffing a large context window; summarise old conversation turns rather than resending them verbatim; drop few-shot examples once the model reliably follows instructions without them.
Sending less also tends to improve accuracy, because models can lose track of detail buried in very long context. It is one of the rare optimisations that helps on both axes.
5. Fix the operational leaks
Retries, timeouts and failed validations all bill normally. So does the traffic your own test suite generates. Separating development from production keys makes both visible, which is usually the first step to reducing them.
Watch for price changes too. Rates move in both directions, and a model you priced against months ago may have been repriced, introductory pricing expiring is a common cause.
Frequently asked questions
What is the fastest way to cut AI API costs?
Route the work that does not need a frontier model to a cheaper one. Because the catalogue spans about 1,600× in blended cost, moving even part of your volume down a tier saves more than every prompt optimisation combined. Test on real traffic to find where quality actually degrades.
Does shortening prompts save much money?
Less than most people expect. Input is the cheaper side of the bill, so trimming it is the fourth-best lever after model choice, output reduction and caching. It is still worth doing (and sending less context often improves accuracy too) but it will not rescue a budget on its own.
Is a cheaper model always worse?
Not for most tasks. Cheaper models are usually smaller or more optimised, which shows on hard reasoning and coding benchmarks, but for classification, extraction, routing and straightforward summarisation the difference is frequently undetectable in production output. The way to find out is to run your own traffic through both.
How much can batching save?
Typically 50% on both input and output rates, in exchange for asynchronous delivery within a stated window. Because there is no quality trade-off, it is close to a free saving for any workload nobody is waiting on, nightly jobs, backfills and evaluation runs especially.
See it in the data
Related guides
Terms used in this guide
Published by Tokenando. Last updated 2026-08-07. Figures in this guide are computed from our own pricing index and dated where they can move; see the methodology and corrections policy.