Token counter

A token is the unit a language model reads and writes. A tokenizer cuts text into pieces from a fixed vocabulary: a common word is often one token, a rare word is split into several, and a space or a punctuation mark can be part of a token. Providers bill and limit requests in tokens, not in words or characters.

For English prose, 1,000 tokens are roughly 600 to 800 words. Google gives 100 tokens ≈ 60 to 80 English words for Gemini. Anthropic says one token is about 3.5 English characters on earlier Claude models, and that current Claude models produce about 30% more tokens for the same text. Code, numbers, URLs, JSON and most languages other than English use more tokens per word.

Each provider trains its own tokenizer with its own vocabulary. A word that is one token in one vocabulary can be two or three tokens in another, so the same prompt has a different count for each model family. The count can also change between model generations of one provider, when the provider changes its tokenizer.

The OpenAI count is exact for the o200k_base encoding: the page runs the same encoding as OpenAI's open-source tiktoken library, in your browser. If OpenAI ships a model with a new encoding, that model's count can differ. Anthropic and Google do not publish a tokenizer for their current models that can run in a browser, so the Claude and Gemini numbers are estimates from the characters-per-token figures that the providers publish. The real count can be higher or lower, most of all for code, numbers and text that is not English.

Use the provider's token counting API with the model you plan to call. Anthropic has a free count_tokens endpoint, and the Gemini API has countTokens. After a call, the usage data in the response gives the exact input and output tokens. Chat requests also add a few tokens per message for roles and formatting, and images, files and tool definitions count separately.

No. The counting runs in a Web Worker on your device. The page sends no request with your text, does not store it, and has no analytics on it. Spell check is off in the text box, because the enhanced spell check of some browsers sends text to a server. To check this, open the Network tab of your browser's developer tools while you type: the only download is the tokenizer data, one time, and it carries no text.

The context window is the maximum number of tokens that a model can take in one request: the system prompt, the conversation, any documents and the answer that the model writes. If the bars show that your text fills most of a window, leave room for the answer, or split the document.

Counting tokens before you send them tells you the size of one prompt. In production, many features call the model many times a day. AI Cost Map wraps your AI client, labels each call with a feature, and shows the token cost of each feature. It sends token counts and labels only: your prompts stay on your servers.