Skip to content

Usage & costs

This page answers two questions an administrator asks regularly: how much are we using AI? and what is it costing us? It covers the whole organization (the Kad chat assistant and every AI agent together), with costs priced from the real prices you recorded on the AI models page. The conversation and token figures used to live at the bottom of the Kad settings page; they now live here, next to the money.

Where to find it

  • Workspace: Admin
  • Path: Admin → Artificial intelligence → Usage & costs
  • Route: /admin/ai-usage
  • Permission: ai_settings:admin

Everything on this page is read-only; it is a report, not a set of controls. Two time windows run throughout: Current month (since the 1st) and Last 30 days (a rolling window). They rarely match, and both are useful: the month for budgets, the rolling window for spotting a change of pace.

Costs

Three cards across the top:

  • Total this month: agents plus Kad, with the last-30-days figure underneath.
  • Agents this month: what the agent fleet actually cost, with the last-30-days figure underneath.
  • Kad this month: what the chat assistant cost. Instead of a 30-day figure, its caption holds the notes described below.

Costs are shown in euros to the cent, and to four decimals when the amount is under a cent. Early on, or with a cheap model, a real total genuinely can be 0.0034 €.

Agent costs are measured. Every model call an agent makes records its input and output tokens, priced there and then at the rates of the model that agent was using, and the result is kept. Tokens read from or written to the provider's cache follow the rule below, the same as for Kad. That is why the AI models prices matter, why a model registered without prices contributes nothing here, and why editing a price later does not rewrite what agents already cost. (When a provider returns no token counts of its own, Kadratis estimates them from the size of the exchange, so a small part of the figure can be approximate.)

Kad costs are measured too. Each Kad answer records the model that produced it, that model's prices at the time and what the answer cost. Switching Kad to another model mid-month or correcting a price later leaves the recorded answers unchanged.

Agent and Kad costs take the provider's prompt cache into account with one rule. A token read from the cache costs the model's Cache read cost when one is entered. Without it, the token costs 0.1 times the input price with Anthropic and the full input price with other providers, so the figure is then an upper bound. A token written to the cache costs 1.25 times the input price with Anthropic and the input price elsewhere.

Answers recorded before Kadratis kept their cost are priced at the rates of the model assigned to Kad today. When the month holds such answers, the card says Partly estimated at the currently assigned model’s rates. When the figure is zero because the assigned model has no price at all (input, output or cache read), it says The assigned model has no cost. Otherwise the card shows no note.

The Kadratis included model costs 0 € by design: it is part of your subscription. An organization running entirely on the included model will see zeros here, and should watch the included-message allowance on the AI models page instead.

Cost by agent and Cost by model

Each table appears underneath as soon as it has a row to show, with a Current month and a Last 30 days column. When only Kad has been used, only Cost by model appears.

  • Cost by agent: one row per agent, so you can see which one is expensive. Pair it with the agent's own Cost per run cap on its Settings tab if a number looks wrong.
  • Cost by model: agent and Kad spend together, sliced by model and sorted with the most expensive over 30 days first. Rows are the provider and model identifiers that were actually called (anthropic:claude-sonnet-5, ollama:mistral), not the friendly names you gave them on the AI models page. When agents and Kad use the same model, they share one row. The Kadratis included model shows at 0. A row labelled Unknown model is older agent activity recorded before per-call model attribution existed.

Cost by agent covers agent runs only. Cost by model leaves out Kad answers recorded before Kadratis kept their cost, so its total can be lower than the cards above.

Conversations

Four unlabelled cards sit between the cost tables and the token table, all about the chat assistant:

  • All conversations: every conversation currently stored. If you set a Conversation retention (days) period on the Kad settings page, conversations eventually get purged and stop counting here.
  • Active conversations (7d) and Active conversations (30d): conversations updated in the last 7 or 30 days.
  • Active users (30d): how many distinct people actually used chat in the last 30 days. The most honest adoption number on the page.

Token usage

One table, two rows (Current month and Last 30 days), with Input tokens, Output tokens, Total tokens, and User messages (the number of questions asked in that window).

This table is about chat, not agents. These are the Kad assistant's tokens: the input count includes tokens served from the provider's cache. Agent consumption is not counted here. It shows up in Cost by agent and in Agent messages below. That is worth remembering when you run on the included model or a local one, where cost is always zero but consumption still is not.

Agent messages (this month)

One card per agent, busiest first. All agents gives the combined count of tickets reviewed this month, then each agent shows its own count, captioned with its last-30-days figure. Every agent you have appears here, including ones that have done nothing yet: a card sitting at 0 is itself worth noticing. Archived agents are left out.

This is the per-organization view of what each agent's own workspace shows individually. Read it alongside Cost by agent: an agent with many messages and little cost is running on a free or cheap model; an agent with few messages and high cost is doing expensive work per ticket and is worth a look.

Tips

  • Compare the two windows, not just the totals. Current month on the 3rd of the month looks tiny; the rolling 30-day figure next to it is what tells you whether anything actually changed.
  • Prices in, costs out. These figures are only as good as the prices on the AI models page. If a cost looks impossibly low, check that the model has prices at all: a model with no price reads as free.
  • Expect a gap with your provider's invoice. The figures use the prices you entered on the AI models page. For a model without a Cache read cost, providers other than Anthropic count cached tokens at the full input price.
  • Use Cost by model when you're weighing a change. It shows what each model is really costing you across your agents and Kad, which is the number to compare before moving work onto a cheaper one.
  • Zero cost is not zero usage. On the included model or a local one, every cost figure stays at €0. Agent messages is where agent load shows up, and the token table is where chat load does.