Providers count tokens to set limits and to bill for usage, and a thousand tokens is roughly 750 English words.

How is text split into tokens?

Each model has a tokenizer that breaks text into pieces it learned during training. Common short words are usually one token, longer or unusual words are split into several, and spaces and punctuation are counted too.

English averages about 1.3 tokens per word, but the ratio varies. Technical terms, names, code, long numbers and many non-English languages use more tokens for the same amount of meaning. You can paste any text into OpenAI's Tokenizer to see exactly where its models split it. Other providers' tokenizers split text a little differently, so the same paragraph gives slightly different counts on different models.

Why do tokens matter for cost?

Usage through an API is billed per token, with separate rates for input (everything you send, including the conversation so far) and output (what the model writes). Output tokens usually cost more than input tokens. A pasted 30-page report therefore costs more than a pasted paragraph, roughly in proportion to its length.

On a prepaid plan, tokens are what your balance is spent on. On a subscription, they sit behind your usage limits. Long inputs also take longer to process, so a very large paste can mean a noticeable pause before the reply starts.

What happens when you reach the token limit?

Every model has a maximum number of tokens it can work with in one go, called its context window, which covers both the input and the reply. Current frontier models accept hundreds of thousands of tokens, and several now take around a million: Anthropic lists 1M-token windows for its recent Claude models (context window documentation), and Google lists just over a million for Gemini 2.5 Pro (Google Cloud documentation). There's also a separate, smaller cap on how long a single reply can be.

What happens at the limit depends on the product:

  • Through an API, a request that's too long is normally rejected with an error.
  • Chat apps usually manage it for you. Some drop the oldest messages, some summarise earlier parts of the conversation, and some ask you to start a new chat.
  • A reply that reaches the output cap stops mid-sentence.

Long before any hard limit, very long inputs can make a model less reliable at recalling details from the middle, so staying under the limit doesn't guarantee everything was taken into account.

An example

You paste a 3,000-word contract and ask for a summary. That's about 4,000 tokens in, plus your instruction, plus a few hundred tokens for the reply, which fits comfortably in any current model's window.

What you can still control is cost and focus. If only the indemnity and termination clauses matter, pasting those two pages costs less and gives the model less to wade through, so its answer on those clauses tends to be sharper.

Common questions

Do I need to count tokens myself?

Rarely. Most products show your usage or a balance and handle the counting. Counting becomes useful when you're estimating costs for a team or building something on an API.

Why did the reply stop halfway through a sentence?

It most likely reached the maximum length allowed for a single reply. Asking the model to continue usually picks up where it left off. If it keeps happening, ask for a shorter answer or split the task into parts.

Why does the same text cost different amounts on different models?

Two things vary between models: how their tokenizer splits the text, and the price per million tokens. A model with a pricier rate or a less efficient tokenizer for your language will cost more for an identical document.