Learn AI /Data and Infrastructure /Cost per token
Cost per token
- Starter
- Data and Infrastructure
- Managers & front offices
- Programme & data teams
Also called: token pricing, usage cost.
Back to the milestone you came from
In five seconds
What it is
The price charged for each small piece of text a model processes or generates, usually counted separately for input and output.
In plain words
It is the price for each small piece of text that goes into or comes out of an AI system.
An everyday example
A city office asks a copy shop to duplicate public feedback forms for a meeting. The shop charges one price for each page you hand over to copy and another price for each page of printed summary notes it prepares for you. If you bring a thick stack of forms or ask for a long summary, the bill rises. Even if two jobs seem similar, the one with more pages costs more.
Pages are uniform and easy to count, but tokens are not the same as words or pages. One sentence can break into different numbers of tokens depending on the language, numbers, formatting and punctuation. Also, a copy shop is not doing the same kind of computation as a model service, so the analogy helps with pricing structure, not with how the system works internally.
Remember thisThe bill usually reflects both the text sent to the model and the text it generates, often at different rates. Longer prompts, longer answers and repeated requests all add cost.
Understand it
Understand it
Why it matters
A non-specialist does not need to calculate every token by hand, but they do need to understand why AI bills rise. Cost per token affects budgeting, procurement, tool choice and workflow design. It helps you ask practical questions: How much text are we sending? How long do we need the answer to be? Which model is affordable for this task? It also helps you avoid false savings. A low price per token can still become expensive if staff send long documents, repeat prompts many times or automate large volumes. Knowing this term supports better planning and fewer surprises.
How token pricing turns a prompt into a bill
After viewing this, the reader should be able to explain that total cost depends on input tokens, output tokens and each rate.
01
1. Send request
The user sends instructions and source text, such as consultation comments to summarise.
Show a box labelled 'Request text' containing three short lines: 'instruction', 'comments', 'context'.
02
2. Count input tokens
The service breaks the request into tokens and counts them for billing.
Show a second box labelled 'Input tokens counted' with a small counter icon and a sample count like '1,200'.
03
3. Generate reply
The model produces an answer, which is also broken into tokens and counted.
Show a third box labelled 'Output tokens counted' with a chat bubble and a sample count like '300'.
04
4. Calculate total cost
Input token count times input rate plus output token count times output rate gives the total usage cost.
Show a final box labelled 'Bill' with a simple formula: '(1,200 × input rate) + (300 × output rate)'.
A left-to-right flow diagram with four boxes: request text, token counting, model response, and bill. Arrows show that the prompt becomes input tokens, the reply becomes output tokens, and both counts feed into the final cost.
Hover or tap each box to reveal a short note: what text is included in that stage and why it affects the bill.
Describe this without the diagram
The diagram is a four-step flow from left to right. Step 1 shows request text, including instructions and source text. Step 2 shows that this text is split into input tokens and counted for billing. Step 3 shows that the model's reply is also split into output tokens and counted. Step 4 shows the billing formula: total cost equals input token count multiplied by the input price, plus output token count multiplied by the output price. The main lesson is that both what you send and what the model returns affect cost.
With reduced motion: A static four-box flow chart shows request text, input token count, output token count and final bill, with the billing formula written in the last box.
Many AI services charge by tokens. A token is a small chunk of text, such as part of a word, a whole word, a number or punctuation. Cost per token tells you how much money the service charges for those chunks.
What goes in is your prompt, system instructions and any extra text you attach. What happens is the service converts that text into tokens, runs the model, and often produces more tokens in the reply. What comes out is a bill based on how many input tokens were sent and how many output tokens were generated.
This matters because two requests that look similar to a person can cost different amounts. A short question with a very long answer can cost more than a longer question with a short answer. Some services also charge different rates for input tokens and output tokens, and some prices vary by model.
Cost per token is about pricing, not model quality by itself. It is also not the same as compute, which means the processing resources used behind the scenes. A cheaper price per token may still lead to a higher total bill if you send large amounts of text or repeat many calls.
How “Paying for photocopied pages” maps onto the real thing
| In the story | In the AI system |
|---|---|
| pages handed to the copy shop | input tokens sent to the model |
| printed summary notes from the shop | output tokens generated by the model |
| price per page | price per token |
| total bill for the job | total usage cost for one or many requests |
A worked example
The officer tests one request to turn cleared background material into a shorter draft statement.
Starting from: The prompt and source passages contain 2,000 input tokens. The requested draft is limited to about 500 output tokens.
- The service counts 2,000 input tokens across the instructions and source passages.
- The model generates 500 output tokens for the draft statement.
- At the illustrative rates provided, input costs 0.50 currency units per 1,000 tokens and output costs 1.50 currency units per 1,000 tokens.
- Input cost is 2,000 ÷ 1,000 × 0.50 = 1.00. Output cost is 500 ÷ 1,000 × 1.50 = 0.75.
- The total charge for this request is 1.75 currency units. Each further revision is metered as another request.
Result: The output is shorter than the material supplied, but it still accounts for a substantial share of the charge because output tokens have the higher illustrative rate. Resending the full source text for several drafting rounds would increase the total.
In a United Nations workplace
The human resources officer supporting an inter-agency secretariat is using an approved AI service to shorten cleared, non-sensitive background material into a statement for delivery. The draft still needs clearance by close of business.
Input and output tokens may be priced differently. Estimating both helps the officer compare prompt options and anticipate the cost of repeated revisions, especially when the full background text is sent again with each request.
Before you try thisDo not enter personnel records, medical information, performance material or identifiable staff comments into a service unless it is specifically approved for that data and task. For this statement, use only the minimum non-sensitive source text required. Human clearance remains necessary.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board reviewing an AI pilot
- A budget discussion before a service subscription is renewed
- A focal point checking usage charges on the provider portal
Avoid confusion
Avoid confusion
Often confused with
Compute
Compute is the processing capacity used to train and run AI systems, with different needs before deployment and during everyday use.
Token
A token is one unit from a model’s vocabulary used to represent input or output, often smaller than a word.
Model hosting
The systems and services that keep an AI model available so it can receive requests, run on computing hardware, and return outputs reliably and securely.
Rate limit
A control that caps how many requests or how much use a service accepts within a set time, to keep systems available, fair and affordable.
Common misconceptions
Common misconception: Cost per token means you pay per word.
Correction: Not exactly. Tokens are chunks of text, and one word can be one token, several tokens or sometimes less than a full token on average across a passage. Numbers, punctuation and formatting also affect the count, so word count is only a rough guide.
Common misconception: A cheap model always costs less to use.
Correction: The listed rate may be lower, but total cost depends on how many input and output tokens your workflow uses. A cheaper model can still lead to a larger bill if prompts are long, answers are long or staff need many retries.
Common misconception: Cost per token tells you how good the answer will be.
Correction: Price and quality are related only loosely. Some higher-priced models may perform better on some tasks, but cost per token is a pricing measure, not a direct measure of accuracy, safety or suitability for your use case.
Quick check
Which situation most directly increases total token-based cost?
With token pricing, the main drivers are the amount of input text and the amount of output text. A longer document adds input tokens, and a detailed reply adds output tokens. The other options do not directly change token counts.
Go deeper
Go deeper
The simple version
It is the price for each small piece of text that goes into or comes out of an AI system.
The technical version
In many language-model services, text is tokenised into units used by the model's processing pipeline. Billing is often metered separately for prompt tokens and completion tokens, sometimes with different rates by model, context window or service tier. Total usage cost is usually calculated from token counts multiplied by the relevant prices, then aggregated across requests. This is a commercial charging metric. It is related to, but distinct from, compute consumption, hosting cost, latency and rate limits.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This official framework helps readers place pricing and usage choices within risk management and governance, and it suits managers, procurement staff and technical teams.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
These official principles help readers connect operational choices such as cost and deployment to trustworthy AI governance, and they suit leadership and policy-aware practitioners.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Cost per token (opens in a new tab)Wikipedia
This search results page helps readers find a range of Wikipedia entries related to the term, and it suits readers who want a broad starting point rather than one definitive article.
search results, not one pagelink checked 2026-07-31
VideoVideo explanations: Cost per token (opens in a new tab)YouTube search
This search results page helps readers find video explainers with different teaching styles, and it suits visual learners who want introductory walkthroughs.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Cost per token (opens in a new tab)Reddit search
This search results page helps readers see common practitioner questions, examples and disagreements, and it suits readers who want informal discussion alongside more formal sources.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
The bill usually reflects both the text sent to the model and the text it generates, often at different rates. Longer prompts, longer answers and repeated requests all add cost.