AI Token Counter

Exact counts need the model's tokenizer vocabulary, which is a large file and differs between providers. The rules used here track English prose closely, typically within about 10 to 15 percent, but they are not a substitute for the provider's tokenizer when a limit is tight.

For ordinary English prose the two rules land close together. The character rule handles text with long or unusual words better, while the word rule suits plain conversational writing. The average smooths out both and is a reasonable middle choice for planning.

Tokenizers are trained mostly on English, so a single CJK character is usually its own token or even split into several. Counting one token per character is closer to reality than dividing by four, though some characters still cost two or three tokens.

Windows range from about 8,000 tokens on older models to 128,000, 200,000 or over a million on recent ones. Check your model's documentation and enter that figure; the counter then shows the share used and the tokens left for the reply.

They count toward the character total, so they affect the character-rule estimate. In real tokenizers a leading space is usually merged into the following word, while punctuation marks often become tokens of their own.

No. The estimate is worked out in your browser from character and word counts, and no model or server sees the text you paste.