# How tokens are counted — Docs

> How euRedact Cloud counts the tokens a document consumes, what each response reports, and every change to the counting method since 14 November 2026, with its dates.

Markdown version of https://euredact.dev/docs/tokens, generated from that page at build on 2026-11-20. Every page of the site is listed with its Markdown version at https://euredact.dev/llms.txt; the full SDK reference is https://euredact.dev/llms-full.txt.

*Documentation*

# How tokens are counted

What a document consumes on euRedact Cloud, what each response tells you, and every change to the counting method.

## What a token measures

euRedact Cloud measures processing in tokens ([General Terms](/legal/terms) 4.4). Your Budget is debited for every model call a request triggers, so the tokens a document consumes depend on:

- the length of the text sent, and the splitting of very long documents into parts;
- the amount and the kind of personal data it contains: a document dense in personal data, or one with special categories of personal data or quasi-identifiers, may undergo an additional pass;
- a fixed instruction overhead per request.

Two documents of the same length may therefore consume different amounts.

euRedact Core, the open-source SDK running on your machine, is never metered.

## What each response reports

Each live cloud response reports what the request was charged and why: `tokens` counts every model call the document triggered, `billing_rate` is 1.0 (a batch is charged at 0.75 and reports its totals on the batch), and `credits` is what was debited, `round(tokens × billing_rate)`. The `factors` name the steps that ran, such as an additional review pass. More personal data also means more output, since each detection is a line the model writes: that counts under `document_length` and has no factor of its own. A batch reports what it consumed as totals for the whole batch, at the batch rate, not per document: on the batch’s status and in the Console, which also shows your consumption overall. Reading it from the SDK: [Python](/docs/python#cloud-result), [Node.js](/docs/nodejs#cloud-result).

*Factor code*

*Meaning*

| --- | --- |
| `rules_only` | Only the rules ran; no model call. |
| `document_length` | The length of the text sent. |
| `instructions` | The fixed instruction overhead per request. |
| `dense_document` | Dense in personal data: a second review pass ran. |
| `special_category_check` | Special-category data or quasi-identifiers were checked; types names which, e.g. MEDICAL\_CONDITION. |
| `long_document` | Very long: retried in two halves. |

A code added later on the service arrives as its string.

## Figures to plan with

- One page is counted as about 2,000 tokens, for presentation only (General Terms 4.4).
- On the 900 synthetic held-out IT, HR and medical documents of our [benchmark](/benchmarks) (about 2,600 characters each, measured 10 November 2026), a document used about 1,450 tokens on average, and a document from the medical set about 1,650. Documents dense in personal data use more; the free plan lets you measure your own.
- Tokens processed through a batch are counted at 75%.
- One document may hold up to 5,000 tokens of text as the gateway counts it on arrival, about 16,000 characters in Dutch or French or 20,000 in English ([instructions for use](/docs/instructions-for-use#limits)). What it consumes, as reported in `tokens`, is more: the instructions and the model’s output count too.

Plans and their Budgets are on [pricing](/pricing).

## Changes to the counting method

euRedact may change how tokens are counted, to improve the quality of detection and the efficiency of the service, and records every change here (General Terms 4.4), starting 14 November 2026. A change that makes the same document consume more tokens takes effect one month or more after it is published here (General Terms 4.4). In practice each change applies to every Workspace on the same date.

1. *Published*

 14 November 2026

 *Takes effect*

 14 November 2026

 *What changes*

 Changelog started; the counting method is as described on this page.
