Methodology & accuracy

What the counter measures

The tool processes the raw text you supply. It does not reconstruct a complete API request, infer hidden instructions, parse images or read your provider account. Each result shows whether it is a local tokenizer count or an offline estimate.

Local counts and model estimates

A local count means the accepted raw text was encoded by the named local engine. It does not prove equivalence to every hosted model, wrapper or billing system. An estimate describes a reference engine or formula whose result can differ from the selected model.

The Claude reference uses a legacy tokenizer, not a verified current Claude tokenizer. Its pieces and IDs belong to that reference. New model prices do not change the counting engine. The same distinction applies to a cost estimate based on an approximate quantity.

Text, files and previews

The counter accepts plain UTF-8 text files up to 5 MB and input up to 1,000,000 UTF-16 code units. It does not extract text from PDF or Word documents. A token preview displays at most 1,200 units; accepted input is counted in full.

Words use language-aware segmentation. Line count and character count describe the supplied text; they are separate from token count. When a token contains part of a multi-byte character, its preview can display that boundary imperfectly. The token ID still identifies the unit.

Browser processing

Counting runs in a browser worker. Tokenizer files load when needed; five-provider comparison is optional in this test version. Empty input does not start vocabulary downloads. Selecting an estimate keeps its warning visible next to the result.

Text is not sent to a provider. The cost calculator accepts numbers and uses local arithmetic. Moving from the counter to the Claude calculator transfers the quantity, model selection and accuracy label through URL parameters. It does not transfer your text.

Sources and updates

The original counter reference review is dated 29 September 2026. Claude prices and model limits on the new pages were checked on 3 October 2026. These dates indicate actual review snapshots; they are not a live feed. Each reference page links to its source.

Use model-specific counting endpoints for a complete request where available, and actual response usage for billing. If you find a discrepancy, share a non-sensitive reproducible example through the project contact page.