Tokens are the small chunks of text that AI language models actually read and write: often a whole short word, sometimes a piece of a longer word, a number, or a punctuation mark. They decide your AI bill because most AI providers charge per token, counting both the text you send in and the text the model writes back, so longer inputs and longer answers cost more.

Once you understand tokens, a lot of AI behavior and pricing stops being mysterious: why pasting a 50-page contract costs more than a quick question, why some languages are pricier than English, and why an AI feature that is cheap in testing can get expensive at scale. This post explains tokens with examples, shows how to estimate usage, and covers the habits that keep costs predictable.

What are tokens in AI, exactly?

A language model does not see letters or words the way you do. Before your text reaches the model, it is chopped into pieces from a fixed vocabulary, and each piece is a token.

Common words are usually one token. Longer or rarer words are split into parts. Spaces and punctuation are often folded in. For illustration, a sentence might break down something like this (the exact split depends on the model):

  • "Please send the invoice" → Please / send / the / invoice
  • "Reconciliation pending" → Rec / onc / iliation / pending
  • "GSTIN 27AAAPL1234C1Z5" → many small fragments, because codes and IDs do not look like ordinary words

A rough rule of thumb for English is that 100 tokens is about 75 words, or a token is around four characters. Treat that as a ballpark only. Different models split text differently, and numbers, code, tables and non-English scripts can use noticeably more tokens for the same amount of meaning.

Why chunks rather than whole words? Because a fixed vocabulary of pieces lets a model handle any word, including names, typos and product codes it has never seen, by building them from parts.

Why do tokens decide your AI bill?

Because tokens are the unit of work. Every token the model reads and every token it writes takes computing power, so providers price their models per token, usually quoted per thousand or per million tokens.

Two details matter:

  • Input tokens are everything you send: your question, plus any instructions, conversation history and documents attached behind the scenes.
  • Output tokens are what the model writes back. They are typically priced higher than input tokens, often several times higher, because generating text is more work than reading it.

Subscriptions like a monthly ChatGPT plan hide this from you. But the moment you build AI into your own workflows through an API, you pay by the token, and design choices start to show up on the invoice.

Why do long documents and long chats cost more?

Because the model rereads everything it needs on every request.

When you chat, each new message is sent along with the earlier conversation so the model has context. Message 20 might carry the tokens of messages 1 to 19 as well. Attach a long PDF and ask five questions about it, and that PDF may be sent with each question.

Request type (illustration only)Roughly what is sentToken weight
Short customer question with brief instructionsA few hundred tokensLight
Reply drafted from a one-page policyA couple of thousand tokensModerate
Question about a 40-page contract, contract sent in fullTens of thousands of tokensHeavy
Long chat on message 30, history includedGrows with every messageCreeps up quietly

The limit on how many tokens a model can consider at once is called the context window, and it is also why AI "forgets" things in very long chats. We explain that in what a context window is.

Do some languages use more tokens than English?

Often, yes. Many models' vocabularies were built mostly from English and other widely written languages, so English text is packed efficiently. Text in Hindi, Tamil, Telugu, Bengali, Marathi and other Indian scripts is frequently split into more, smaller pieces, so the same message can cost noticeably more tokens. The gap varies by model and has been narrowing in some newer ones, so test with your own real messages rather than assuming.

This matters for businesses serving customers on WhatsApp in regional languages. A support agent that answers in Hindi may cost more per conversation than the same agent in English. That does not make it a bad idea; it just belongs in your estimate. Our guide to multilingual AI agents for Indian languages covers quality as well as cost.

How do you estimate token usage for a business workflow?

You do not need to be exact. You need the right order of magnitude before you commit. A simple method:

  1. Pick the workflow, for example "answer incoming WhatsApp questions from our product catalog".
  2. Estimate input per request: instructions plus the relevant catalog snippet plus the customer's message plus recent history. Paste a realistic example into a provider's token counter to get a real number.
  3. Estimate output per request: how long is a typical answer? A three-sentence reply is small; a full draft proposal is large.
  4. Multiply by volume: requests per day times working days per month.
  5. Apply the provider's current per-token prices for the model you are considering, and add a safety margin for retries and longer-than-average cases.

For illustration: if each conversation uses a few thousand tokens and you handle a few hundred conversations a day, monthly usage lands in the tens of millions of tokens. Whether that is cheap or expensive depends entirely on which model you choose, which is why model choice matters so much.

How do you keep token costs under control?

Most savings come from sending less and choosing well, not from haggling.

  • Send only what is needed. Instead of attaching a whole manual, retrieve the two relevant sections. This is what retrieval-augmented generation (RAG) does.
  • Trim conversation history. Summarize older messages rather than resending them in full.
  • Ask for shorter outputs where length adds nothing. Output tokens are the pricier ones.
  • Use the smallest model that does the job. Sorting emails does not need the most capable model; drafting a tricky legal summary might. See small language models for business.
  • Cache repeated content. Many providers offer discounts when the same long instructions are reused across requests.
  • Watch usage from day one. Set budgets and alerts so a looping automation cannot run up a surprise bill.

Our guide to LLM cost optimization goes deeper on each of these.

Frequently Asked Questions

How many words is 1,000 tokens?

For ordinary English text, roughly 750 words, though it varies by model and content. Numbers, code, tables and non-English languages usually use more tokens per word, so measure with your own text when cost matters.

Are tokens the same across ChatGPT, Claude and Gemini?

No. Each model family uses its own way of splitting text, so the same paragraph can be a different number of tokens on different models. Compare costs using each provider's own token counter on your real content.

Why are output tokens more expensive than input tokens?

Generating text is more computationally demanding than reading it, because the model produces output one token at a time. Most providers price output higher to reflect that, which is why asking for concise answers is an easy saving.

Do I pay for tokens on a ChatGPT subscription?

Not directly. Consumer and team subscriptions charge a flat fee with usage limits. Token billing applies when you use AI models through an API, which is how custom business tools and automations are built.

Do images and audio use tokens too?

Yes, in models that accept them. Images, audio and video are converted into tokens of their own, and larger or longer files generally use more. Check the provider's documentation for how each media type is counted.

Want an AI workflow with a predictable bill?

If you are planning an AI assistant or automation and want a realistic cost estimate before building, tell us about the workflow. We will size the token usage, suggest the right model for each step, and design it to stay efficient as volume grows. See our agentic AI development work.