AI token counter for GPT, Claude, Gemini and more

Language models don’t read letters or words. They read tokens: pieces of text from a fixed vocabulary, usually whole common words, parts of rarer words, punctuation and spaces. API prices, context windows and output limits are all measured in tokens, so knowing how many tokens a prompt or document uses tells you what it costs and whether it fits.

Every model family has its own tokenizer

A tokenizer is built by learning which pieces of text appear most often in the training data. Different companies use different data and vocabulary sizes, so the same text produces different counts. For plain English the differences are small, but code, numbers and non-English text can vary a lot: see the examples below. That is why this tool counts each model separately instead of using one rule of thumb.

Where the counts come from

OpenAI models use the tokenizers OpenAI publishes in its tiktoken library. The tool runs the same encodings in your browser through the open-source gpt-tokenizer library, picks the encoding using OpenAI’s own model-to-encoding mapping, and is tested to give identical results to tiktoken. OpenAI hasn’t published which tokenizer the GPT-6 family uses, so those counts use o200k_base, the tokenizer of GPT-4o through GPT-5.x, and are labelled “Unconfirmed”.

Open-weight models such as DeepSeek, Qwen and Mistral publish their tokenizer files. When you pick such a model, the tool downloads its tokenizer from this site, pinned to an exact version and checked against the publisher’s hash, then runs them with Hugging Face’s tokenizers library. They are tested against the official implementation on thousands of samples in many languages, with code and emoji. Where a model’s tokenizer isn’t available here yet, the tool says so instead of guessing.

Claude and Gemini tokenizers aren’t public. Rather than show an estimate dressed up as a count, the tool shows OpenAI’s count clearly marked “for reference only” and offers the provider’s own counting API, which is free, with your API key. Your key and text go directly from your browser to Anthropic or Google. Anthropic describes its count as an estimate that can differ slightly from a real request.

Tokens, words and cost

As a rough guide for English, one token is about four characters or three-quarters of a word, so 1,000 words is around 1,300 tokens. Each row also shows what the text would cost as input (in a prompt) and as output (if a model wrote it), at the model’s current list price. Output usually costs several times more. For a whole workload with daily volumes, caching and batch pricing, use the LLM API cost calculator.

The count covers only the text you enter. A real request adds a few tokens per message for the chat format, plus any system prompt, tool definitions and images. If you are checking whether a long document fits a model, leave room for those and for the reply.