Why token counts differ between models

Paste the same text into two tools and you may get two numbers. Both can be right.

Different tokenizers

Each model family has its own vocabulary and its own way of splitting text. The Gemini documentation describes its tokenization in its own terms, and Anthropic's in its own. So the count for one model is not the count for another, and the 4-characters rule is an average for English, not a law.

A tokenizer can change

Anthropic's glossary says that its newer models (Claude 4.7 and later, and Claude Mythos Preview, as the page said on 14 November 2026) use a newer tokenizer that produces approximately 30 percent more tokens for the same text than older models. So the same text can cost more tokens after a model upgrade even though the words did not change. Budgets should be checked again when a model changes.

Counts are estimates even from the maker

Anthropic's token counting page says the token count is an estimate, that the actual number of input tokens used when creating a message might differ by a small amount, and that the count may include tokens added automatically by Anthropic for system optimizations, which are not billed. Your bill reflects your content.

What to do

1. Use the maker's counter for the model you will really use.
2. Treat the estimator on this site as a first look, with a range.
3. Add a margin of safety to budgets, and recheck after model changes.

This is an independent guide. It is not connected with Google, Anthropic, OpenAI or any company named here. Facts were read on 14 November 2026 from the pages named on each page. The numbers are estimates from rules of thumb, not the counts of any maker's tokenizer.