AI Token Counter & LLM Cost Calculator.

A token counter is a tool that tells you how many tokens a piece of text becomes once a language model reads it. Models are billed, limited and rate-limited in tokens rather than words, so that count is what decides whether a prompt fits and what it costs. This token counter runs the real tokenizer in your browser across 34 models from 7 providers and shows the count, the cost and the share of the context window together.

A token is the basic unit of text an AI model processes. It can be a whole word, a fragment of one, a punctuation mark or a single character. Common English words are usually one token each, while longer or rarer words split into pieces — "tokenization" arrives at the model as "token" plus "ization". For ordinary English, one token averages about 4 characters, or 0.75 of a word.

About 4 characters of English prose, which works out to roughly 0.75 of a word. That is an average, not a rule: "the" is one token of three characters, while a long URL or a base64 blob can cost a token per character. In money, one input token on GPT-5.6 Terra is $0.000002 — which is why rates are always quoted per million rather than per token.

To tokenize text is to split it into the vocabulary units a model actually reads, then map each unit to an integer ID. The model never sees your characters; it sees that list of IDs. Tokenizing is deterministic — the same text always produces the same tokens — and reversible, so decoding the IDs returns the original text byte for byte. Counting tokens means tokenizing the text and reporting how many pieces came out.

The five-type classification comes from compiler theory rather than AI: a lexer sorts source code into keywords, identifiers, operators, separators and literals. Language models do not use those categories. For LLM tokenization the useful split is different — whole words, subword fragments, single characters or raw bytes, punctuation and whitespace, and special control tokens such as end-of-text markers. Only that last kind is added by the API rather than by your text.

About 750 words of English prose, using the working ratio of 0.75 words per token. Going the other way, 1,000 words is roughly 1,333 tokens. Both directions are estimates and only hold for prose: 1,000 tokens of minified JSON, source code, or Hindi and Japanese text carries far fewer words than that. Counting beats dividing for exactly that reason.

Around 150,000 words, or about 300 printed pages at 500 words a page — a full-length novel. It is also a common context-window size, so "does this fit in 200K" is usually the real question. As spend, 200,000 input tokens cost $0.400 on GPT-5.6 Terra and $0.030 on GPT-4o Mini.

About 750,000 words, or roughly 1,500 pages of printed text — call it eight average novels. Several models now accept that much in a single request: GPT-5.6 Sol has a 1M-token window and Claude Sonnet 5 has 1M. Reading a million tokens aloud would take somewhere near ninety hours.

Roughly 1,500,000 words, or about 3,000 printed pages. No model in the picker takes that in one request — the largest window here is 1M tokens — so two million usually describes a monthly budget or a document set you will need to chunk. As spend it is $4.00 of input on GPT-5.6 Terra, or $0.300 on GPT-4o Mini.

"1M tokens" is shorthand for one million tokens, and it appears in two places. As a context window it means the model can hold about 750,000 words of prompt, documents and history at once. As a pricing unit it is how every provider quotes rates — $2.00 per 1M input tokens on GPT-5.6 Terra works out to $0.000002 per token. It has nothing to do with crypto tokens.

On GPT-5.6 Terra, at $2.00 per million, 1,000 input tokens cost $0.00200 and 1,000 output tokens cost $0.012 — output is always the expensive direction. Across the whole picker the same 1,000 input tokens range from $0.00015 on GPT-4o Mini to $0.010 on Claude Fable 5, a 67× spread. The text does not change; only the bill does.

One million tokens is exactly one billing unit — every provider quotes rates per million, so this is the number printed on the pricing page. On GPT-5.6 Terra a million input tokens is $2.00, which means one dollar buys 500,000 tokens, or about 375,000 words. Output is dearer at $12.00 per million. At the cheap end GPT-4o Mini charges $0.150, so the same dollar goes 13 times further.

Token count multiplied by the selected model's published rate, with input and output priced separately. Paste or upload text here and it is counted as input, with the per-call cost shown immediately. Cached input — where consecutive requests share a prompt prefix — is cheaper again: $0.200 per million on GPT-5.6 Terra against $2.00 standard. Rates come from each provider's own pricing page and are re-checked regularly.

Three ways, in descending order of effort. Paste the text into the counter at the top of this page and read the number — exact for OpenAI models, because the real o200k_base vocabulary runs in your browser. Call a tokenizer library yourself: tiktoken for OpenAI, SentencePiece for Google, Hugging Face tokenizers for open models. Or divide your character count by 4 for a rough guess that drifts badly on code, JSON and non-Latin scripts.

For a quick answer, paste text or drop a file into the counter above and the count appears as you type — no account, no API key, nothing uploaded. In code, use the provider's own tokenizer: len(enc.encode(text)) with tiktoken in Python, or js-tiktoken in JavaScript. Where a vocabulary is not published, use the provider's counting endpoint instead — Anthropic's count-tokens endpoint is the only exact way to get a Claude figure.

By running a byte-pair-encoding pass, not by counting characters or words. The tokenizer splits your text with a fixed regex, then repeatedly merges the most frequent adjacent byte pair according to a merge table learned during training, until no merge applies. The pieces left over are the tokens, each mapped to an integer ID. o200k_base, the encoding current OpenAI models use, holds about 200,000 such pieces; the older cl100k_base holds about 100,000.

For OpenAI models, exactly — it runs OpenAI's own open-source BPE vocabulary in your browser, so the number matches what you are billed. 11 of the 56 models in the comparison are counted that way. Anthropic, Google, xAI, DeepSeek and Meta do not publish their tokenizers, so those counts are calibrated estimates and are labelled as estimates next to the total. Treat them as a close guide, not a billing figure.

Yes. PDF, DOCX, TXT, Markdown, JSON, CSV and XML are supported, up to 10MB, and several files at once is fine — each gets its own line and the totals add up. PDFs are parsed entirely in your browser with PDF.js; nothing is uploaded. A scanned PDF with no embedded text cannot be read, and the tool says so rather than returning a misleading count.

It converts text into the integer IDs a model can process, and converts them back again. Concretely: it splits the text into candidate chunks with a regex, runs byte-pair merges over each chunk against a fixed vocabulary, and emits one ID per surviving piece. It also handles the awkward parts — the leading space that makes "hello" and " hello" two different tokens, bytes that are not valid UTF-8, and the control tokens that mark where a message begins and ends.

Every text operation a model performs begins and ends with one, so the tokenizer is what makes your input legible to the model at all. Outside the model you use one to count tokens before sending a request, to price a prompt, to check a document against a context window, to chunk text for a vector database on token boundaries rather than character ones, and to trim conversation history so it still fits.

Tokenization is splitting text into the units a model reads. "Token counter" is two tokens, because both words are common enough to have their own vocabulary entry — and the space belongs to the second one, so " counter" and "counter" are different tokens. The word "tokenization" is rarer, so it splits into "token" plus "ization": one word, two tokens. Rarer still breaks down further — "unbelievably" costs three, as "un" plus "bel" plus "ievably". The visualiser above shows the exact split for any text you paste.

In AI, the model's own vendor. OpenAI publishes tiktoken, Google open-sourced SentencePiece, Meta ships its tokenizer with the Llama weights, and Hugging Face's tokenizers library implements BPE, WordPiece and Unigram for open models. Anthropic, Google's Gemini and xAI keep their production vocabularies closed and offer counting endpoints instead. If you came looking for payment tokenization, that is an unrelated service from card networks and payment processors.

Pick an algorithm — BPE, WordPiece or Unigram — assemble a corpus that resembles the text you will actually process, choose a vocabulary size, and run a trainer such as Hugging Face tokenizers or SentencePiece; it outputs a vocabulary and a merge table. A larger vocabulary means fewer tokens per document but a bigger embedding matrix. One caveat matters most: you cannot retrain a tokenizer for an existing model, because its vocabulary is welded to the embedding weights.

Yes, completely free and with no limits — no sign-up, no API key, no trial, no paid tier holding the useful part back. Everything runs in your browser, so there is no server to meter and nothing for us to bill: you can count 1,000,000 tokens as easily as ten. The 34 models in the picker are all available to everyone, and so is file upload.

No. The tokenizer, the file parsing and the arithmetic all run locally in the page you are already looking at. Nothing is ever sent to a server, because there is no server to send it to — which is what makes this usable on a draft contract, an unreleased spec or a customer transcript. Close the tab and nothing persists.