Token counter
Free, instant, no signup. Paste any text — a prompt, a system message, a document, a chat log — and get its token count, its characters and words, and what that text costs as input or as output on every Claude and GPT model, each next to its provider's list price. Counting runs in your browser.
- tokens
- 0
- characters
- 0
- words
- 0
Counted in your browser. The tokenizer downloads once, the first time you enter text, and nothing you paste is sent anywhere.
What this text costs, per request
Paste text above to see how many tokens it is and what it costs on each model.
Exact for GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.5: gpt-tokenizer maps these models to o200k_base, the tokenizer counted here.
≈ marks an estimate: the o200k_base count of the same text. GPT-6 is newer than the tokenizer's model list, so its tokenizer is not confirmed there, and Claude tokenizes with its own vocabulary.
Anthropic does not publish Claude's tokenizer. Its count_tokens endpoint is free and returns Claude's own number, and Anthropic's docs say Claude 4.7 and later models produce about 30% more tokens than earlier Claude models for the same text.
相同文字在其他語言中需要更多 token
Kunavo 的指南以十三種語言發布，因此同一份內容可以計算十三次。在 266 份指南中，涵蓋 13,501 個段落和 586,465 個英文單字，以下是各語言使用的 o200k_base token 數，並以英文為 1：
| 語言 | 相較於英文的 Token 數 | 每個 Token 的字元數 |
|---|---|---|
| English | 1.00× | 4.62 |
| Chinese (Simplified) | 1.18× | 1.76 |
| Indonesian | 1.21× | 4.41 |
| Portuguese (Brazil) | 1.22× | 4.38 |
| Spanish | 1.23× | 4.52 |
| German | 1.29× | 4.39 |
| French | 1.31× | 4.41 |
| Italian | 1.35× | 4.07 |
| Vietnamese | 1.38× | 3.72 |
| Chinese (Traditional) | 1.39× | 1.50 |
| Korean | 1.41× | 2.05 |
| Turkish | 1.44× | 3.62 |
| Japanese | 1.64× | 1.64 |
英文散文平均每個單字為 1.30 個 token，因此 1,000 個單字約為 1,300 個 token，而一百萬個 token 約可容納 770,000 個單字。以英文提示詞估算價格時，同一項產品在日文輸入 token 上的費用約高出 64%，在韓文則約高出 41%。
閱讀這項資料有兩個限制。翻譯採用 Kunavo 自行提供的版本，因此翻譯較精簡或較冗長，都會讓該筆資料移動。這些數字是使用 OpenAI 的 o200k_base 計算的：Claude 的分詞器會以自己的方式切分文字，因此對 Claude 而言，倍數不會相同。若要將數量換算成每月帳單，請使用 AI token 費用計算器。
Price what you counted
Price a workload per month
Cache hits and writes priced
Cached input and reasoning
Every Claude rate, vs list
GPT rates by model
Every model, vs official
How many tokens is 1,000 words?
About 1,300 tokens for English prose with o200k_base, the tokenizer of OpenAI's GPT-4o and GPT-5 models: across 586,465 words of Kunavo's own guides the average was 1.30 tokens per word. Code, numbers, URLs and most other languages take more tokens per word, so paste your own text above for the real figure.
How many words is 1 million tokens?
Roughly 770,000 English words, at the 1.30 tokens per word measured on Kunavo's guides with o200k_base. A million tokens holds fewer words in other languages: the same guides took 1.64 times as many tokens in Japanese, 1.29 times in German and 1.18 times in Simplified Chinese.
Is the token count exact for Claude?
No — for Claude it is an estimate. Anthropic does not publish Claude's tokenizer, so the figure shown for Claude models is the o200k_base count of the same text. Anthropic's token counting endpoint, /v1/messages/count_tokens, is free and returns Claude's own number, and Anthropic's documentation says Claude 4.7 and later models produce about 30% more tokens than earlier Claude models for the same text.
Which models is the count exact for?
GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.5: gpt-tokenizer, the open-source library that does the counting here, maps those models to o200k_base, the encoding it counts with. GPT-6 is newer than that library's model list, so its count is marked as an estimate rather than assumed. Either way the figure is the text itself — a chat request adds a few formatting tokens per message.
Is the text I paste uploaded anywhere?
No. The tokenizer runs in your browser: its vocabulary downloads once, the first time you enter text, and after that every count happens on your device. The text is not sent to Kunavo or to anyone else, and it is gone when you close the page.
Why does my API usage show more input tokens than this counter?
Because a request carries more than your text. Chat formatting adds a few tokens per message, the system prompt and tool definitions are sent on every call, images are billed as tokens of their own, and reasoning models bill hidden reasoning as output. The usage field of each API response is the number you are billed for; this counter is for sizing a prompt before you send it.