Token counter for AI prompts

How to use the token counter

1. Type or paste your text in the box, or pick one of the examples.
2. Read the two counts: the o200k count is exact for that tokenizer; the Claude count is an estimate (see below).
3. Look at the words: blue ones cost one token, yellow two, orange three, pink four or more. Long, rare or foreign words, numbers and dates split into several tokens.
4. Compare: paste the same sentence in another language, or a vague request next to a precise one.

How the tokenizer works (no AI model is called)

The counter uses o200k, an open tokenizer (the one OpenAI publishes for its recent models), loaded into the page the first time you use the box. It cuts your text into tokens exactly as that tokenizer does, on your own device.

Claude's tokenizer is not public, so the Claude figure is an estimate. We measured it once, on 36 texts of our own in English, German and French (emails, prompts, meeting notes, reports, tables, code), by counting each with Claude's own token count and with o200k. Claude counted, per o200k token:

The counter guesses your text's language from its commonest words and applies that ratio (1.77, the ratio over all 36 texts, when it cannot tell). Tables of figures sit at the low end of the range, ordinary prose at the high end: the estimate can be off by about 20% on a given text.

What the result means

A token is the unit an AI model reads, writes and bills: often a whole short English word, a piece of a longer one, a digit group or a punctuation mark. The same message costs more tokens in German or French than in English, and a date or an invoice number costs more than its length suggests.

The cost table shows what reading your text once costs at each model's list price for input tokens, and 1,000 times, as a team sending it every day would.

The last block is the part most people miss. An AI agent does not read your request alone: at every step it takes (reading a file, making a change, checking the result) it reads its own instructions again. The agent of our training sessions re-reads about 20,700 tokens of instructions at every step, so a single step costs at least about 21,000 tokens, and a request that takes 2 to 5 actions 45,000 to 117,000. A clear request that needs fewer steps saves far more than a shorter one.

FAQ

Is this token counter free?

Yes, and it needs no account. It runs entirely in your browser.

Which tokenizer does the token counter use?

o200k, the open tokenizer OpenAI publishes for its recent models (a GPT tokenizer). Its count is exact for that tokenizer; the Claude figure is the estimate described above.

Is my text sent anywhere?

No. The tokenizer is a file the page loads once; counting happens on your device. No AI model is called and nothing you type leaves the page.

Why does Claude count more tokens than o200k?

Each model family has its own tokenizer, with its own vocabulary. On our measured texts, Claude made about 1.5 tokens for every o200k token in English and about 2 in German. That is why the same prompt can cost differently from one model to another, beyond the price per token.

How accurate is the Claude token counter estimate?

It is an average measured on 36 texts, so it is close on ordinary prose and less so on tables, code or unusual text. For an exact count, the model's own API has a token counting endpoint.

Why do German and French texts make more tokens?

On our 36 measured texts, the same message took more tokens in German and French than in English: Claude counted about 2.04 tokens per o200k token in German and 1.74 in French, against 1.47 in English. Paste a long German compound word into the counter and you will see it split into several pieces.

How do I spend fewer tokens with an AI agent?

Say in your first request what you want, with the context, the task, any guidelines and the constraints: an agent that has to ask, guess or redo a change re-reads its instructions at every extra step. In our sessions, "improve it" took 113,000 tokens where the precise request took 49,000. Our training lets each employee practise this with a real agent: book a demo to see a session.