There are not very many ways to reduce an AI bill. There are maybe a dozen, they are all known, and most teams reach for them in close to the worst possible order — starting with the one that is easiest to talk about rather than the one that moves the most.
This is the full list, ranked by leverage against effort. Your account will reorder it. That is fine; the point is to start from the whole set rather than the first idea in the room.
Tier 1 — usually large, usually cheap to do
Stop making the call. The largest reduction available is always the request you do not send. Cached identical requests, features nobody uses, a model called on every keystroke instead of on submit, an enrichment step whose output is discarded downstream. This sounds glib and it is routinely the biggest single line. Look for calls whose results nothing reads.
Cut the prompt that grew. System prompts accumulate instructions and never shed them. This is pure prefill, it is billed on every single call, and the fix is reading it and deleting. Cheapest real money on this list.
Turn on prompt caching, correctly. Where a provider offers it, a stable prefix re-reads at a large discount — the price index shows published rates by model. The correctness caveat is load-bearing: caches match on prefix, so the stable part must come first. Put the user's variable input ahead of the system prompt and you get nothing while believing you are covered.
Right-size retrieval. Fetching 20 chunks when the answer lives in the top 3 is a 6x input bill on that step. Measure where in the retrieved context answers actually come from before assuming more is safer.
Tier 2 — usually large, real work
Cap and shape output. Output bills at multiples of input — see the output premium. Length limits, structured responses instead of prose, short field names, and instructions that actually shorten answers all bite directly on the expensive side. Requires quality testing, which is why it is Tier 2 and not Tier 1.
Route by difficulty. Not every request needs your best model. A classifier or a cheap first pass that escalates only hard cases is the highest-ceiling lever on this list — and the one most likely to quietly degrade quality if you skip evaluation. Build the eval before the router, not after.
Question the reasoning model. Extended reasoning bills as output. It earns its cost on genuinely hard tasks and is a large silent waste on easy ones. This is a per-call-site question, not a per-application one.
Batch the asynchronous work. Where a provider offers a batch or off-peak tier and the work is not user-facing, the discount is typically substantial for latency you were not using anyway. Underused because nobody remembers it exists.
Tier 3 — real, but smaller or slower
Switch models. The lever everyone starts with. It is real, and the price index exists to make it easy — but it is capped by the price spread between comparable models, while the levers above are capped by how wasteful your architecture is, which is usually the larger number. Do it, but do not do it first.
Negotiate. At meaningful volume, list price is not the only price. Requires spend that justifies the conversation.
Move the workload. Self-hosting, an alternative provider, a smaller fine-tuned model. Occasionally correct, frequently a large project justified by a spreadsheet that omits the engineering time. Be honest about the fully loaded cost.
Fine-tune to shorten prompts. Trading a long few-shot prompt for a tuned model can pay off at high volume. It also creates a model you now own and must maintain. Genuine trade, not a free win.
The ordering that actually matters
Rank your own list by (expected saving) ÷ (engineering days), and be honest about the denominator. A 40 percent saving on a step that is 2 percent of the bill is a rounding error dressed as a win. A 10 percent saving on the step that is 60 percent of the bill is the quarter's work.
Two failure modes worth naming, because they are the common ones:
Optimizing what is measurable instead of what is large. Teams fix the thing they have a dashboard for. Get the attribution from reading your first bill first, or you will do this too.
Shipping a saving that costs more than it saves. Three engineer-weeks to save two hundred dollars a month is a bad trade, and it is easy to make when the saving is a percentage and the cost is time. Do the arithmetic in dollars, both sides.
What none of this is
There is no lever on this list that is free, and none that works without someone measuring afterward to confirm quality held. Anyone selling a switch that cuts your bill in half without changing anything about what you send or what you get back is selling something else.