What Are Tokens in AI? The Meter Behind Your AI Bill
A plain-English explanation of tokens, why input and output are priced differently, and how to work out what AI will actually cost your firm each month.
A token is a small chunk of text that an AI model reads and writes. One token is roughly three-quarters of a word, so 1,000 words works out to about 1,300 tokens.
Tokens matter because they are the unit AI vendors bill you in. Every word you send and every word the model sends back is counted and priced.
If you pay a flat monthly fee for a chat plan, you never see this meter. The moment your firm connects AI to a document, a mailbox, or a workflow, the meter becomes your bill.
This guide explains what a token is, how tokens are counted, why output costs more than input, and how to estimate a real monthly figure from a real task.
What is a token in AI?
A token is a small piece of text — usually a whole short word, or part of a longer one. AI models do not read letters or sentences. They break text into tokens and process those.
The useful rule of thumb is that one token is about three-quarters of an English word. So 100 words is roughly 130 tokens, and 1,000 words is roughly 1,300 tokens.
Common words like "the" or "and" are one token each. A long or unusual word like "indemnification" gets split into several.
- 1 token ≈ 3/4 of an English word
- 1,000 words ≈ 1,300 tokens
- One page of a business document ≈ 500 words ≈ 650 tokens
- Punctuation, spaces, and line breaks are counted too
Paying a flat monthly rate for AI and wondering what per-token billing would actually cost your firm? We will count the tokens on your real work and give you a monthly number in 30 minutes.
Book a ConsultationHow are AI tokens calculated?
Tokens are calculated by the model's tokenizer, a fixed piece of software that splits your text into pieces before the model sees it. You do not choose the split, and it is the same every time for the same text.
Every request has two counts. Input tokens are everything you send: your question, any document you paste, and the earlier turns of the conversation. Output tokens are everything the model writes back.
Both counts are added to your bill on every single request. That last point surprises people most.
In a long chat, the whole conversation so far is re-sent as input each time you hit send. Turn 20 costs far more than turn 2, even if you typed the same short question.
- Input tokens = your prompt + pasted documents + the prior conversation
- Output tokens = everything the model writes back
- Different model families split text slightly differently, so counts vary a little
- OpenAI publishes a free tokenizer page where you can paste text and see the exact count
Why input and output tokens are priced differently
Output tokens cost more than input tokens — usually five to six times more. Vendors price them separately because they cost very different amounts to produce.
Reading your text is fast. The model processes the whole input at once. Writing a reply is slow, because the model generates one token at a time, and each one depends on the one before it.
The published rates show the gap plainly. Claude Sonnet 5 is $2 per million input tokens and $10 per million output tokens. Claude Haiku 4.5 is $1 in and $5 out.
One caveat on that Sonnet 5 rate: $2 / $10 is introductory pricing, and it is scheduled to rise to $3 / $15 on September 1, 2026, so check the vendor's current pricing page before you budget from it.
The practical takeaway is simple. Pasting a long document into a prompt is cheap. Asking for a long document back is not.
- Input is processed in one pass, so it is cheap
- Output is generated one token at a time, so it costs more
- Claude Sonnet 5: $2 per million input, $10 per million output (Anthropic)
- Claude Haiku 4.5: $1 per million input, $5 per million output (Anthropic)
- Cheapest habit to build: ask for short answers unless you need long ones
Why your chat subscription never shows a token count
A flat monthly subscription hides the meter on purpose. Plans like ChatGPT Plus and Claude Pro cost $20 per user per month, and that price does not move with how much you type.
The meter is still running underneath. The vendor absorbs it and controls their cost with usage caps instead — the message limits that stop you mid-session on a busy day.
So a subscription is predictable, and that is worth real money to a small firm. You know the number before the month starts.
The trade is that you cannot connect a subscription to your own systems. Anything automated runs through an API key, which is billed per token with no flat fee. Our guide on an AI API key vs a subscription covers that choice in full.
- Subscription: fixed price, hidden meter, hard usage caps
- API key: no fixed price, visible meter, no caps at your volume
- A seat covers one human typing; an API key covers software calling the model
- Some vendors sell prepaid credits instead — see how AI credits work
The two billing shapes tokens are sold in
Tokens reach you in one of two shapes: a flat monthly seat, where the vendor decides your token budget in advance and hides the meter, or a per-token API bill, where you pay for exactly the tokens you use.
The difference that matters for understanding your bill is predictability. A seat is a fixed number known before the month starts; an API bill moves with usage and is only knowable after the fact.
Many firms end up with both. Seats for the partners, an API key for the one workflow that runs on its own.
Choosing between them is a separate decision with its own trade-offs on caps, volume, and cost. Our guide to an AI API key vs a subscription works through it properly.
- Predictability — Seat: fixed, known before the month starts. API: varies with usage.
- Who it serves — Seat: one named human. API: any script, tool, or workflow.
- Where the meter shows — Seat: hidden from you. API: itemised per request.
How much does one real task cost? A worked example
One realistic professional task costs a few cents. Here is the arithmetic, using a job a small advisory or law firm actually does.
The task: summarise a 20-page client document and draft a two-page memo from it.
Step one, count the input. Twenty pages is about 10,000 words, which is about 13,000 tokens. Add your instructions and round up to 14,000 input tokens.
Step two, count the output. A two-page memo is about 1,000 words, or roughly 1,300 output tokens.
Step three, apply the rates. On Claude Sonnet 5, input is $2 per million and output is $10 per million (Anthropic).
Input: 14,000 ÷ 1,000,000 × $2 = $0.028. Output: 1,300 ÷ 1,000,000 × $10 = $0.013. Total: about 4 cents.
Run the same task on a flagship model and it costs more. On GPT-5.6 Sol, at $5 in and $30 out per million (OpenAI), the same job costs about 11 cents.
- Input: 14,000 tokens × $2 per million = $0.028
- Output: 1,300 tokens × $10 per million = $0.013
- Total per memo on Claude Sonnet 5: about $0.04
- Total per memo on GPT-5.6 Sol: about $0.11
- Do this 40 times a month: about $1.64, or about $4.36 on the flagship
How much will AI tokens cost me a month?
For a two-person professional firm doing normal document work, per-token AI costs land in single-digit dollars a month. It only climbs when software starts making the requests instead of a person.
A human can only type so much. Even a heavy user struggles to spend $10 a month in tokens by hand, which is why a $20 seat is usually the more expensive option for light users.
Automation removes that ceiling. A workflow that processes every inbound email or every uploaded file makes thousands of requests without anyone noticing.
When we run our own scheduled content and SEO routines across the sites in our portfolio, every run bills per token against an API key. The consistent pattern is that the bill tracks how much those routines write, not how many of them fire.
The estimate you want is one line of arithmetic: tokens per task × runs per month × the published rate. Our AI model cost calculator does it for you, and AI model pricing lists the current rate for every major model.
- One person, occasional document work: often under $5/month on an API key
- One person, all-day chat use: a $20 seat is simpler and caps the risk
- An automated workflow: cost scales with runs, so estimate before you switch it on
- Set a hard spend limit in the vendor's billing console on day one
- Formula: tokens per task × runs per month × published rate
When token costs start to matter for a small firm
Token costs matter the moment a process runs without a person watching it. Until then, the numbers are too small to manage.
Three things push a small firm past that line: connecting AI to a mailbox or document store, running a task on a schedule, and feeding whole files in rather than short questions.
Any one of those turns a cents-per-task habit into a monthly line item worth checking. If yours has crossed that line, our guide to AI token cost optimization covers the tactics that bring the bill down.
- You connect AI to email, a shared drive, or your case management system
- A task runs on a schedule instead of when someone clicks
- You start sending whole documents rather than short questions
- More than one person's work is going through the same API key
Frequently Asked Questions
- Tokens are the small chunks of text an AI model reads and writes. One token is roughly three-quarters of an English word, so 1,000 words is about 1,300 tokens. AI vendors bill by the token, counting both what you send and what the model sends back.
- Every request is split into input tokens and output tokens. Input is your prompt plus any pasted text plus the earlier conversation. Output is the model's reply. Both are counted on every request, which is why long chat threads cost more per message as they grow.
- Rates are quoted per million tokens and vary by model. Claude Sonnet 5 is $2 per million input tokens and $10 per million output tokens, and Claude Haiku 4.5 is $1 and $5 (Anthropic). The Sonnet 5 figure is introductory pricing scheduled to rise to $3 / $15 on September 1, 2026, so check the vendor's current pricing before budgeting. A typical document-summary task costs a few cents. See our AI model pricing page for current rates across every major model.
- Because writing is slower than reading for a model. Input is processed in a single pass, while output is generated one token at a time. That gap is why output usually costs five to six times what input costs.
- Not directly. A $20-per-month seat is a flat fee, so you never see a token count. The vendor controls their cost with usage caps instead, which is why you get cut off during heavy sessions. Token billing only becomes visible when you use an API key.
- A tokenizer splits your text into pieces before the model reads it. The split is fixed, so the same text always produces the same count. Common short words are one token, and long or unusual words are split into several. OpenAI publishes a free tokenizer page where you can paste text and see the exact number.
- It is the same thing as an AI token. LLM stands for large language model, the type of AI behind chat assistants. An LLM token is one chunk of text the model processes, roughly three-quarters of a word.
- About 650 tokens. A typical business page holds around 500 words, and 1,000 words is roughly 1,300 tokens. A 20-page document is therefore about 13,000 tokens of input.
- Multiply tokens per task by runs per month by the published rate. Count the input and output separately, since output is priced higher. Our AI model cost calculator runs this for you against current rates.
Not sure what AI would actually cost your firm each month?
Book a free 30-minute AI workflow audit with Layer3 Labs. We will count the tokens on your real tasks, compare a flat seat against per-token billing, and give you a monthly figure you can plan around.
Book Your Free AI Workflow Audit