Token counter

Paste a prompt, a document or any text, and the count appears with the characters, the words and the characters per token. Tokenizer picks the vocabulary: o200k for GPT-4o, GPT-4.1, GPT-5 and the o3-mini and o4-mini models, cl100k for GPT-4, GPT-3.5 Turbo and the text-embedding-3 models. Show switches between the count, each token on its own, and the token IDs the model actually receives.

From also offers JSON to Token cost by format. Paste JSON and the same data is written as indented JSON, minified JSON, YAML, CSV and XML, each counted, so you can see which format fits the most data into a prompt.

Language models don't read letters or words. They read tokens, chunks of text from a fixed vocabulary of about 100,000 entries in cl100k and 200,000 in o200k. A common word is usually one token, often with the space before it attached, so " world" is a single token. Rarer words split into pieces, numbers split into groups of up to three digits, and text in other scripts tends to take more tokens per character than English.

Tokens matter because they're what you pay for and what fills a model's context window. A prompt's cost and whether it fits are both measured in tokens, not characters.

For a list of records with the same fields, CSV is usually cheapest, because it writes each key once in the header instead of in every record. Minified JSON drops the indentation that indented JSON spends tokens on, and YAML drops most of the quotes and braces but still repeats every key. XML repeats each key twice, in the opening and closing tag. The comparison runs on your own data, since nesting, long strings and numbers all shift the balance.

Each model family has its own tokenizer, and only some are published. The counts here are exact for OpenAI's models and only a rough guide for anyone else's. Anthropic offers a free token counting endpoint for Claude that returns its own estimate, and notes that Claude Opus 4.7 and later use a newer tokenizer that turns the same text into roughly 30 percent more tokens than earlier Claude models. That endpoint needs an API key and receives your text, so it isn't used here.

The tokenizer runs in your browser. The first time you count, the page downloads the vocabulary for the tokenizer you've picked, a file of up to about 1 MB from this site, and it's cached after that. Nothing you type is sent anywhere.