# Koins

> How inference is metered and what each request, image and transcription costs.

Koins are the balance a workspace spends on AI work. Every message the agents answer, every image and every transcription costs koins, and everyone in the workspace (members, agents, routines and background work) spends from the same balance.

## Getting koins

- **Starting koins.** Each account gets 50,000 koins once, on the first workspace it creates. Later workspaces start at zero, and deleting and recreating a workspace doesn't give them back. Joining someone else's workspace doesn't add koins.
- **Plans.** Subscribing to a plan adds its koins every month. See [Billing](/docs/billing.md).

Koins don't expire and don't reset monthly. What you don't spend carries over.

## What things cost

These are the default prices.

| What | Cost |
| --- | --- |
| A message | 1 koin per 125 tokens of input and output, at least 8 koins |
| A routine run | Twice the normal price |
| An image | 800 koins |
| An audio transcription | 8 koins |
| An embedding (turning text into a vector for search) | 1 koin per 125 tokens, at least 8 koins per request |

A token is roughly four characters of text. A short exchange usually costs the minimum; long documents, long conversations and multi-step tool work cost more. A request is charged once, even when Koris tries several models or several steps to answer it.

Some things cost nothing:

- a request that fails before any AI work starts, or that a provider refuses;
- a request you cancel before it reaches the model;
- website-watch checks that find no change.

The **Billing** page shows a pricing table built from your installation's actual prices.

## Seeing your balance

The sidebar shows your remaining koins, with a bar that turns amber when a fifth of your latest grant is left and red at zero. The **Billing** page shows:

- available koins;
- koins **reserved** for work that is still running;
- koins spent;
- the history of every charged request, with routines labelled by name.

## Running out

Before a request starts, Koris sets aside the most it could cost. If the workspace doesn't have enough koins for that, the request doesn't run. When it finishes, you're charged what it actually used, never more than was set aside, and the rest goes back to your balance.

When the balance reaches zero, agents stop answering until koins are added. Routines and background work stop too.