How many Claude tokens is my text?

Claude’s tokenizer isn’t published, so no web page can count Claude tokens exactly without asking Anthropic’s API. Anthropic’s glossary gives the rule of thumb: a token is about 3.5 English characters. The counter divides the size of your text by 3.5 and shows a range around it, from ÷ 4.5 to ÷ 2.5, because code, digits and most languages other than English take more tokens per character.

It measures the size in bytes (UTF-8), not in characters: an accented letter takes two bytes and a Chinese, Japanese or Korean character three, which raises the estimate for the scripts that usually take more tokens per character. It is the rule the UsageMeter extension uses to weigh a file before you attach it on claude.ai, so both give the same figure.

A new Claude generation can come with a new tokenizer. Anthropic says Claude 4.7 and later models, the Fable models included, use a newer one: the same text produces about 30 % more tokens than on earlier models, and the exact increase depends on the content. The same prompt can therefore count more tokens on Claude Opus 5.5 than on Claude Sonnet 4.6.

For an exact figure, Anthropic’s API has a token counting endpoint: free, within rate limits, it counts a message with the tokenizer of the model you name. In Claude Code, the Session block of /usage shows the tokens a session actually used.

Images are counted exactly, by the rule in Anthropic’s vision documentation: Claude first scales an image down to its model’s limits (1,568 pixels and 1,568 visual tokens, or 2,576 pixels and 4,784 tokens on Claude 4.7 and later models), then counts one token per 28 × 28-pixel patch. A 1920 × 1080 screenshot is 1,560 tokens on Claude Sonnet 4.6 and 2,691 on Claude Opus 5.5.

PDFs are read page by page, as text and as an image of the page. Anthropic puts the text at typically 1,500 to 3,000 tokens per page, and the page image is counted like any image. The counter reads the page count from the file itself and gives a wide range per page, from a light page to a dense one; for a PDF it cannot see into, it says so rather than guess.

On Pro, Max and the other plans, Anthropic publishes limits as percentages of a 5-hour session and of a week, never as a number of tokens. That is why this counter never turns tokens into a share of your limit. What it does show is what weighs: long messages, attachments and the length of the conversation all count, because the whole conversation goes along with every new message.

Settings › Usage in Claude shows the share of each limit you have used. The UsageMeter extension keeps those gauges on screen, adds the tokens of the current session with a rough estimate of what is left, and shows the weight of a file before you send it.