Work out token count estimate instantly with clear inputs, formula shown and shareable results.
Byte-pair encoders split English into pieces averaging about four characters, equivalently around 1.33 tokens per word. Estimating from both characters and words and averaging the two is more robust than either alone, because character counts overestimate for text full of short words and word counts underestimate for code, URLs and non-English scripts.
Token heuristics
tokens ~ characters / 4; tokens ~ words / 0.75; blended = average of both estimates
For English prose it is usually within 10 percent. For code, JSON, or languages like Thai and Japanese it can be off by two to three times, so tokenise exactly when cost matters.
Yes. Vocabulary size and merge rules differ per model family, so the same text can cost noticeably different token counts across providers.