Token Counter

One text, four different numbers

A token is not a word and not a character. It is a chunk of bytes that one model learned to treat as a single unit, and each family learned a different set of chunks. The word antidisestablishmentarianism is six tokens on GPT-5. A waving hand emoji is two there and three on GPT-4 Turbo. A paragraph of Russian costs about twice what the same paragraph costs in English.

So this page asks you to pick a model before it prints a number. A counter that shows one figure for everything is showing you one tokenizer and hoping you do not notice.

Where the number is exact

OpenAI publishes its tokenizers and Meta publishes the Llama 3 one, so those counts are the real thing. They run here in your browser and they match tiktoken, checked against OpenAI's own test fixtures every time the tests run. The badge says Exact and names the encoding: o200k_base for GPT-4o and later, cl100k_base for GPT-4 and GPT-3.5.

Anthropic and Google publish nothing you can run offline. For those the badge says Estimate and the page shows a range, built from the ratios each vendor documents itself. I would rather show a range than a number I cannot back. Claude Opus 4.7 changed tokenizer and the same text now costs about thirty percent more, so the older Claude models here use the older ratio instead of one average for both.

The split is where the money is

The tiles under the box are the real token boundaries. Long words break into pieces. A space usually rides along with the word after it. A tile ringed in amber is one character that took several tokens. That is the expensive case, and it matters most in a prompt you send a thousand times a day.

The line table ranks what to cut. A pasted log or a base64 blob will often cost more than every instruction around it, and the share column says by how much.

What the count leaves out

The context window holds the whole conversation, not one message. Your system prompt, the tool definitions, every earlier turn, the images and the model's own reply all draw on the same budget. Reasoning models spend hidden tokens on thinking that no counter can see.

So read the bar as a floor. If it says a prompt fills eighty percent of the window, you have less room left than it looks. If you need the billed figure to the token, every vendor has a count endpoint that returns it, and every one of them wants your API key.

Drop the file instead of pasting it

The question behind most of these searches is whether a document fits in one request. I checked thirty token counters for this. Not one of them has a file input. Every one is a bare text box, so you paste a contract in and hope the clipboard held all of it.

This page takes the file. PDF, DOCX, Markdown, JSON, CSV or a source file, dropped on the box or picked from the dialog. The reading happens here, the same way the counting does. A PDF pulls the pdf.js worker from a public CDN the first time, and your file still never leaves the tab.

You do not sign in. The vocabulary file is a few megabytes and arrives the first time you pick a model that needs it, with the count showing a spinner until it lands. After that it sits in the browser cache.

If you are counting because a prompt got too long, the Word and Character Counter gives you the platform limits instead, and the JSON Formatter will shrink a payload before you paste it into a prompt.