AI token costs: how to cut your AI bill without cutting corners

What a token is, in plain English

AI models don't read words, they read tokens. A token is a chunk of text, usually a short word or part of a longer one. As a rough guide, 1,000 tokens is about 750 words of English.

When you use an AI model through its API, or through a tool that's built on one, you pay for the tokens you send in (your instructions, documents and chat history) and the tokens the model sends back. Output tokens usually cost several times more than input tokens.

Why AI bills creep up

Most surprise bills aren't caused by one big job. They come from small habits repeated thousands of times: a long set of instructions sent with every request, a whole document pasted in when one page would do, or a chat history that keeps growing until every reply re-reads the whole conversation.

The other common cause is using the most powerful model for everything. The top models are excellent, but sorting emails or tagging invoices doesn't need them.

Seven ways to cut token costs

1. 01Trim your instructions. Long system prompts are often full of repeats and examples that no longer help. Cutting them in half is common.
2. 02Send less context. Search for the relevant section of a document instead of sending the whole thing.
3. 03Cache what repeats. Prompt caching lets you pay a lot less for the parts of a request that stay the same each time.
4. 04Use the right model for each job. Route simple tasks to a smaller, cheaper model and keep the big model for work that needs it.
5. 05Ask for shorter answers. Output tokens cost the most, so ask for the format you need and nothing more.
6. 06Batch work that can wait. Overnight reports and bulk jobs can usually go through a batch service at a lower rate.
7. 07Set budgets and alerts so a runaway script can't run for days unnoticed.

Keep quality where it was

Cheaper is only better if the results hold up. Before changing anything, save a set of real examples with answers you're happy with. After each change, run the same examples again and compare. If quality drops, put that part back.

This is where a person who knows the work matters. AI can suggest a shorter prompt, but someone who knows what a good invoice coding or customer reply looks like has to judge whether it's still right.

Where to start

Export a month of usage from your AI accounts and sort it by task. In most businesses a handful of jobs make up most of the spend. Start there, make one change at a time and measure the saving in pounds.

What is a token in AI?

AI tokens are the units a large language model reads and writes. Before the model processes your text, it breaks it into tokens and turns each one into numerical representations it can work with. When it replies, the model generates one token at a time, predicting the next token based on everything before it.

A single token can be a whole short word, part of a longer word, a space or a punctuation mark. Common words are often one token. Rarer words, names and numbers are often split into two or more tokens. That's why the same text can give different token counts on different models.

How AI breaks text into tokens

Most providers use a method called byte pair encoding, or a close relative of it. It learns which chunks of text appear most often and gives each common chunk its own token. English tends to need fewer tokens than many other languages, and code or tables can need more tokens than plain prose.

You don't need to understand the maths. The practical point is simple: more text in or out means more tokens, and more tokens cost money.

How many tokens is a page of text?

As a rough guide, a thousand tokens is about 750 English words. A typical email might be a few hundred tokens. A long report can run to tens of thousands. Most providers have a free tool that shows how many tokens a piece of text uses, which is worth trying with your own documents.

How AI token costs are worked out

Most AI services that you use through an API have token based pricing. Providers charge for input and output separately, usually as a price per million input tokens and per million output tokens. Each AI model has different prices, and the bigger, more capable frontier models cost much more than budget models.

Token pricing changes often as new models arrive, so always check the provider's current page. In this guide we use [current rate per million tokens] rather than numbers that may be out of date by the time you read it.

Some models also produce reasoning tokens. These are the model's working before it gives you its answer. You don't always see them, but they're usually billed as output tokens, so a model that thinks for longer on complex queries can cost more per answer.

The hidden tokens in every request

A single API call sends more than the question you typed. It usually includes the system prompt with your instructions, any documents or search results you've added, the conversation history so far and, for an AI agent, the tool definitions that describe what it can do. All of that counts as input tokens, every time.

Factors that influence your AI costs

A few factors influence most AI token costs. If you know which ones apply to you, you know where to look first.

• Which model you use: expensive models can cost many times more per token than cheaper models
• How much context you send: more context per request means more input tokens
• How long the answers are: output tokens usually cost the most
• How long conversations run: the whole history is re-sent with each message
• How many steps an AI agent takes: each step is another API call
• How often jobs run: a small cost per request adds up across thousands of requests

How to reduce AI costs with cheaper models and model routing

Frontier models are excellent at hard work like drafting a careful reply or analysing a contract. Many simple tasks don't need them. Sorting emails, tagging invoices or pulling a date out of a document can usually go to a smaller, faster model at a fraction of the token cost.

Model routing means sending each job to the smallest model that does it well. Some teams route by task type. Others try a cheaper model first and only pass the request to a bigger one when the answer isn't good enough. Either way, test with real examples before switching, because the underlying model matters for quality as well as cost.

LLM token usage in AI agents and automations

An AI agent that books meetings, updates your CRM or answers customer questions may make several calls to the model for one task. Each call sends the instructions, tool definitions and history again. Token consumption can grow quickly, and most companies only notice when the invoice arrives.

Keep an eye on the maximum number of steps an agent can take, and set rate limits so a loop can't run unchecked. Shorter instructions, fewer tools per agent and summarised history all help. Prompt caching also has a significant impact here, because the instructions and tool definitions are the same on every call.

Cut AI API costs and keep cost predictability

Cutting costs is half the job. The other half is knowing what next month will look like. Set a budget and alerts in each provider's billing systems, give each tool its own API key so you can see who is spending, and review usage monthly.

Be careful with free tokens and free tiers. They're useful for testing, but limits and terms change, and data on some free tiers may be used differently from paid plans. For business AI applications, a paid account with clear limits gives you more control.

Finally, measure cost per outcome, not just cost per token. If an assistant answers a customer question for less than a penny, you can compare that with the time it saves. That's where you find out which AI workloads give you the most value, and where more tokens are worth paying for.