How to Reduce LLM Token Costs (Without Wrecking Quality)

The cheapest way to reduce LLM costs is to stop paying for tokens that do no work: a system prompt padded with dead instructions, retrieved documents the answer never uses, chat history re-sent long past the point it helps, and answers longer than anyone reads. Cut those, cache what repeats, and only then route easy tasks to a cheaper model. Downgrading every call to a weaker model is the blunt lever, and pulling it first can cost more in rework than it saves.

Find the bigger half first

Every call has an input side and an output side, and they are priced differently. On Anthropic's price sheet for Claude Sonnet 5.5, as of October 2026, output costs $10 per million tokens and input $2. Divide one by the other and you get the break-even: input costs more than output only when a call sends more than five input tokens for every one it gets back.

Pull a week of usage and compute that ratio per feature. A document classifier sending 8,000 tokens to get a 50-token label sits far above five, so input is where the money goes. A drafting tool sending 500 tokens to get 2,000 back sits far below, so output is. Chat products drift upward over a conversation as history piles onto the input side. The ratio tells you which cuts below to try first. Your own provider's rates give you your own break-even, so redo the division for each model you run.

Seven cuts, from least to most risky

1. Cache what does not change. As of October 2026, Anthropic bills Sonnet 5.5 cache reads at $0.20 per million tokens, a tenth of the input rate, and says a 5-minute cache pays for itself after one read. A 10,000-token reference document sent on 30,000 calls a month is $600 of input uncached and $60 at the cache-read rate, plus the writes.
2. Batch what can wait. Anthropic's Batch API takes 50% off both input and output for requests that do not need an answer right away, such as overnight reports or bulk tagging.
3. Trim the system prompt. Read it as if you had never seen it and cut every instruction that no longer solves a real problem. Prompts grow one edge case at a time and rarely shrink on their own.
4. Tighten retrieved context. If you ground answers in documents, send fewer, better-matched chunks instead of a wide net "just in case." Every extra chunk is billed as input on every call.
5. Cap conversation history. Decide how many earlier turns the model needs, and summarize or drop the rest instead of re-sending the whole thread forever.
6. Constrain response length and format. Ask for a set length, a structured format or a direct answer. Output is the pricier side, so this cut works on the expensive half.
7. Route by task difficulty. Send simple classification, extraction and formatting to a smaller, cheaper model, and keep your most capable model for work that needs its reasoning.

The trap: cutting quality to chase a number

The fastest way to damage an AI feature is to force shorter, vaguer answers or move every call to the cheapest model, whatever the task. Users can notice a worse product before finance notices a smaller bill, and the rework can cost more than the tokens did. Our post on the true cost of an AI agent shows the math. In its illustrative example, raising the share of accepted results from 80% to 90% lowers cost per accepted result more than cutting token cost by a third.

The cheaper model is not always cheaper per task, either. A model with a lower rate can use more tokens to finish the same job, and Anthropic's own docs warn that its newer tokenizer counts about 30% more tokens for the same text. The opposite also happens. Anthropic prices Claude Sonnet 5.5 the same as Sonnet 5 and says, from its own testing, that it costs "up to 30% less per task than its predecessor" and typically needs far fewer tokens to do the same work. Neither claim was measured on your work. The only way to know which way a swap cuts is to run it on your own tasks and count.

Measure before and after

A cost cut is real only if you can show both numbers. Use the same method twice: calls per month, average input and output tokens per call, and current rates, once before the change and once after. Count tokens on a sample of real prompts rather than estimating. Anthropic's token-counting endpoint is free to use, subject to rate limits. The main token-cost guide has a calculator for a quick before-and-after check.

Write down the date and the rate you used for each side. A saving measured at one price can shrink when the price moves, and Google has already said Gemini 4 Argon's introductory $2 and $10 per million will rise to $4 and $20.

When cost cutting is the wrong tool

Sometimes the number that needs fixing is not the per-call cost. The task may not be worth automating at any price, and trimming its tokens will not change that. Our post on the AI automation threshold scores whether one workflow is worth it at all. And no amount of trimming protects you from a runaway loop or a leaked key. LLM API cost control covers the controls that cap spend, with a hard cutoff as the backstop.

Pairs well with

Every cut that changes a prompt or a model needs a test, and the Prompt Regression Lab ($89) snapshots a baseline, diffs each prompt change and returns ship, hold or regressed. The AI Cost-Per-Task Calculator ($29) answers the earlier question, whether a task is worth automating at all, with Automate, Review or Skip. Once your cuts are live, the AI Spend Runaway & Billing-Safeguard Gate ($49) grades whether a sudden spike would be stopped.