AI Token Counter and GPT Tokenizer

How to use the token counter

Paste or type your prompt, or open or drop a plain-text file (.txt, .md, .csv, .json or source code). The big number is the exact token count for the tokenizer you pick, and the table compares the other models. Under See the tokens you can view the text as coloured tokens, as a token list, or as the raw token IDs, which you can copy or download as JSON.

What a token is

Language models read tokens, not letters or words. OpenAI models use byte-level byte pair encoding (BPE). The text is first cut into chunks by a regular expression: a word together with the space before it, up to three digits, a run of punctuation, or whitespace. Each chunk becomes UTF-8 bytes, and neighbouring bytes are merged, most frequent pair first, until only pieces from the vocabulary are left. A common English word is one token including its leading space, while a rare word, a long number or most non-English text splits into several. An emoji or a Chinese character can take two or three tokens, and a token can even hold only part of a character’s bytes; those pieces are shown in a dashed box.

Which tokenizer each model uses

Counts here match OpenAI’s tiktoken library token for token. Like its encode_ordinary, special markers such as <|endoftext|> typed into the text count as ordinary text. The open-weight gpt-oss models use o200k_harmony, which only adds special tokens to o200k_base, so plain text gives the same count.

Claude and Gemini estimates

Anthropic and Google don’t publish tokenizers for their current models that can run in a browser, so this page estimates from the character count using each company’s rule of thumb: about 3.5 characters per token for Claude and about 4 for Gemini. Anthropic says the tokenizer introduced with Claude Opus 4.7 can use up to 1.35 times as many tokens as earlier Claude models for the same text, so the Claude estimate is a range. The rules describe English prose. Code, numbers and non-Latin scripts need more tokens per character, and the page warns you when most of your text is not in the Latin alphabet.

For an exact figure, send the text to Anthropic’s POST /v1/messages/count_tokens endpoint or the Gemini API’s countTokens method, naming the model you plan to use.

Why an API request uses more tokens

• Chat requests wrap each message in role markers, which adds a few tokens per message, plus the system prompt and any tool definitions.
• Images, audio and files are counted separately.
• Output tokens, and reasoning tokens on reasoning models, are billed on top of the input.

Tips for fewer tokens

• Numbers split into groups of up to three digits, so long IDs, hashes and base64 strings are expensive. Leave them out when the model doesn’t need them.
• Pretty-printed JSON and wide Markdown tables cost more than the same facts in plain sentences.
• Watch Characters per token. English prose usually lands around 4 to 5 with these tokenizers, and a much lower number points to text that is costly to send.