> For the complete documentation index, see [llms.txt](https://docs.artific.nl/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.artific.nl/en/for-administrators/reports/tokens.md).

# Token report

What your usage costs, and where it goes. Found under **Reports → Tokens**.

## What a token is

Models measure work in tokens, roughly a word-part. Everything is counted in two directions:

| Direction  | What it is                                                                                                                       |
| ---------- | -------------------------------------------------------------------------------------------------------------------------------- |
| **Input**  | Everything sent to the model: the question, the instructions, the conversation so far, and any documents retrieved to answer it. |
| **Output** | What the model wrote back.                                                                                                       |

Input is usually the larger share, and the part people find surprising. A short question against a large knowledge base sends a great deal of context.

## The screen

The same organisation, period and granularity controls as the [Activity report](/en/for-administrators/reports/activity.md).

### Summary cards

| Card               | What it shows                                            |
| ------------------ | -------------------------------------------------------- |
| **Total cost**     | Total token cost for the period and scope.               |
| **Input / output** | The split between the two, in both cost and token count. |

### Chart

Choose **Token cost** or **Token usage**, then how the bars are stacked:

* **Input / output**: where the cost falls between question and answer.
* **Model mix**: how much each model contributes.
* **Source split**: how much comes from assistants against the Toolbox.

### Filters and breakdown

Filter by provider or by model. Break the table down **Per organization** (where you manage children), **Per user**, **Per assistant**, **Per toolbox element**, **Per model** or **Per provider**.

Breaking down per model adds unit-cost columns, which is how you compare what each model actually costs you rather than what it costs per token in the abstract.

## Using it to control cost

**Find the expensive assistant.** Break down per assistant. Cost is rarely spread evenly. One assistant is usually most of the bill.

**Look at the input share.** A high input share means large amounts of context are being sent. Two settings control that directly: **Assistant context size** in [advanced settings](/en/for-administrators/assistants/advanced-settings.md), and how many results a search tool returns. Both trade cost against how much the assistant can see.

**Match the model to the job.** A high-end model answering simple, high-volume questions is the most common avoidable cost. Compare per-model unit costs, then move straightforward assistants to a faster model and check [quality control](/en/for-administrators/quality.md) still passes.

**Check the trend, not the day.** Cost per active user over months tells you whether usage is becoming more efficient. A single busy day tells you nothing.

**Watch for surprises.** A sudden jump usually has a specific cause: an assistant went live, a search tool started returning far more context, or an integration started retrying.

{% hint style="info" %}
This report can be switched off for an organisation. If it is missing, or you see a message that token insights are disabled, ask an organisation administrator (or, for a child organisation, the parent) to enable it.
{% endhint %}