What Is a Token in AI? How AI Pricing Really Works
A token is the unit AI APIs bill you for — roughly four characters of text. See how tokens are counted, why output costs more than input, and how to estimate and cut your AI bill.
On this page
In AI, a token is the basic chunk of text a language model reads and writes. It is roughly four characters, or about three quarters of a word. AI APIs charge per token. They bill the input (your prompt) and the output (the reply) separately, and output tokens usually cost several times more than input tokens.
If you build or run anything on an AI API, tokens are the unit you pay for. Knowing how they work is the difference between a predictable bill and a nasty surprise. This guide explains what a token is, how providers turn tokens into a bill, and how to estimate and cut what you spend. You can check any figure here against our live AI pricing comparison.
What Is a Token in AI?
A token is the smallest unit of text a large language model (LLM) processes. It is not quite a word, and not quite a character. It is a "subword" fragment that the model's tokeniser has learned to recognise. Common words are usually a single token. Longer or rarer words get split into several pieces.
According to OpenAI's documentation, one token is about four characters of English, or roughly 0.75 words. As a working rule of thumb:
- 1 token is about 4 characters of English text.
- 100 tokens is about 75 words.
- 1,000 tokens is about 750 words, or roughly a page and a half.
Tokens vs Words vs Characters
Models are billed in tokens, not words or characters, because tokens are the real unit of work. Every token takes computing power to process, whether it sits in your prompt or in the reply. Charging per token simply reflects that cost.
The split is not always intuitive. The word "darkness", for example, is often broken into two tokens, "dark" and "ness". Numbers, code, emoji and punctuation can each use their own tokens too. That is why a short snippet of code can cost more tokens than a longer line of plain prose.
How Text Becomes Tokens: Tokenisation Explained
Tokenisation is the process of breaking text into tokens before a model can read it. As NVIDIA puts it, tokens are the language and currency of modern AI. Every prompt and every response is turned into tokens, processed as numbers, then turned back into text.
Here is the catch. Each provider uses its own tokeniser, so the same sentence can produce different token counts on different models. The identical request can therefore cost slightly different amounts on GPT, Claude and Gemini, before you even compare headline prices. To dig into the mechanics, Hugging Face has a clear primer on tokenisers.
How AI API Pricing Actually Works
AI APIs use token based pricing. You are charged for the number of tokens processed, almost always quoted as a price per million tokens and billed in US dollars. There is no extra per message or per request fee, just the tokens themselves. Each model usually shows two numbers, because input and output are priced separately.
Input Tokens vs Output Tokens
Providers split every request into two kinds of token:
- Input tokens (also called prompt tokens): Everything you send to the model. This covers your instructions, any context, system prompts and attached documents.
- Output tokens (also called completion tokens): Everything the model generates in reply.
Both are counted and added to your bill. A long document pasted into a prompt drives up the input cost. A long generated answer drives up the output cost.
Why Output Tokens Cost More
Output tokens almost always cost more than input tokens. The gap is usually two to six times, with a median of around four times. The reason is technical. Input tokens are processed in a single "prefill" pass that can run in parallel. Output tokens are made one at a time, and each new token needs a full pass through the model.
So generating text is slower and more demanding than reading it, and the price reflects that. The practical lesson is simple. Long replies can quietly inflate a bill even when your prompts are short. Tools that produce a lot of text, such as summarisers, need tight limits on output length.
Context Windows, Caching and Other Cost Factors
Token prices are not the whole story. A few other levers shape what you actually pay:
- Context window: This is the maximum number of tokens a model can consider at once. Bigger contexts let you send more, but more input means a higher cost, and some providers raise the rate past a size threshold.
- Prompt caching: Several providers bill repeated, unchanged input at a fraction of the standard rate, often around 10 percent. That is a big saving for repetitive prompts.
- Batch processing: Many providers offer roughly half price rates for non urgent work that can be returned within a set window, such as 24 hours.
How Much Do AI Tokens Cost? A Provider Comparison
Prices vary enormously, from a few cents to tens of dollars per million tokens. The figure depends on the model's size and capability. The table below shows representative rates as of mid 2026. Providers update prices often, so always confirm current figures and see all 10 providers in our live comparison.
| Provider and Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| OpenAI GPT-5.5 | $5.00 | $30.00 |
| OpenAI GPT-5.4 | $2.50 | $15.00 |
| Anthropic Claude Sonnet 4.6 | $3.00 | $15.00 |
| Google Gemini 2.5 Pro | $1.25 | $10.00 |
| xAI Grok 4.3 | $1.25 | $2.50 |
| DeepSeek V4 Flash | $0.14 | $0.28 |
The spread is the real headline. A budget model can be more than a hundred times cheaper per token than a flagship. For many jobs, a mid tier model does the work at a fraction of the cost of the largest option.
How to Estimate Your AI API Bill (With a Worked Example)
The formula for any single request is simple:
Cost = (input tokens / 1,000,000 x input price) + (output tokens / 1,000,000 x output price)
Say you want to summarise 1,000 customer support emails a day. You pick a model priced at $2.50 per million input tokens and $15.00 per million output tokens. Assume each email is about 500 input tokens, and each summary is about 200 output tokens. Here is how the cost breaks down:
- Input: 1,000 emails x 500 tokens = 500,000 tokens. At $2.50 per million, that is about $1.25.
- Output: 1,000 summaries x 200 tokens = 200,000 tokens. At $15.00 per million, that is about $3.00.
- Daily total: Roughly $4.25, or about $128 a month.
Notice that output is the bigger share, even though it is the smaller token count. That is the higher output rate at work. To run the numbers for your own model and volumes, use our interactive cost calculator.
How to Reduce Your Token Costs
Once the mechanics are clear, several savings become obvious:
- Trim your prompts: Cut redundant instructions and context. Every token you remove is a token you do not pay for, on every call.
- Cap output length: Set a sensible maximum on generated tokens so replies cannot run away with your budget.
- Use prompt caching: If you reuse the same system prompt or context, caching can bill it at a fraction of the standard rate.
- Batch non urgent work: Move background jobs to discounted batch processing where a short delay is fine.
- Right size the model: Save flagship models for hard tasks, and route simple, high volume work to cheaper ones.
API Pricing vs ChatGPT Plus: What Is the Difference?
This trips up a lot of people. A consumer subscription such as ChatGPT Plus is a flat monthly fee for the chat app. It has usage limits, but no per token charge.
The API is a separate, pay as you go product. You are billed per token for every request your software makes, with no fixed monthly cost. One is a product you log in to use. The other is infrastructure you build on.
For anything programmatic, you are in API territory. That is exactly where token pricing applies.
Frequently Asked Questions
How Many Words Is 1,000 Tokens?
About 750 words of English, since one token averages roughly 0.75 words. The exact figure depends on the text, because code, numbers and non English characters use more tokens.
Are Input and Output Tokens Priced the Same?
No. Output tokens almost always cost more than input tokens, often two to six times as much. Generating text is more demanding than reading it.
Why Do Different Models Charge Different Amounts for the Same Text?
There are two reasons. First, each provider sets its own per token price based on model size and capability. Second, each uses its own tokeniser, so the same text can count as a different number of tokens.
Is the API Cheaper Than a ChatGPT Subscription?
It depends on usage. For light, occasional use, the API can cost pennies. For heavy daily volume, costs scale with tokens and can pass a flat subscription. Compare them against your real workload, not headline prices.
How Can I Count Tokens Before Sending a Request?
Most providers publish tokeniser tools that count tokens for a given piece of text. You can use them to estimate cost before you send anything. As a quick guide, divide your character count by four.
Fun Fact: The Language You Prompt In Changes Your Bill
Because every provider uses its own tokeniser, the same sentence can ring up a different bill on GPT, Claude and Gemini. It goes further still. Writing in languages such as Arabic, Chinese or Hindi can use 20 to 30 percent more tokens than the same meaning in English. In short, the language you prompt in literally changes how much you pay.
The Bottom Line
Tokens are the unit that turns AI from a clever demo into a line on your invoice. A token is about four characters. Input and output are priced separately, and output is the expensive half. Once you know that, AI costs stop being mysterious and become something you can plan around.
The single biggest variable is which model you choose, where prices differ by orders of magnitude. So before you commit to a provider, compare the live rates and run your own numbers with the compareai.today calculator. A few minutes of estimation now can save a great deal at the end of the month.
