What an agent's cost is made of and how to count it
We have no public price list. The three parts that make up the cost of an agent, which of them you see yourself, and how to tell whether the agent pays off.
Three parts of the cost
The model — on your account
A Claude or ChatGPT subscription, a key from another provider or Amazon Bedrock in your AWS is in your name and paid by you at the provider’s published rates. The hub does not resell compute and adds no markup. Which input to choose is your decision.
Infrastructure — at cost
One deployment on AWS — in the region you choose or in your own AWS account. You see the real bill, with no multiplier on top.
The hub’s work — per deployment
We work out the price once the task, the channels and the number of people involved are clear. It is named before the work starts and does not change because you added a second agent: a new agent is configuration, not another seat in a licence.
Why there is no price list
What the first agent costs is determined by its work. A rate invented before we knew the task would be wrong both ways: too expensive for a simple role, too cheap for a complex one.
If you need a number before a conversation, say so in the form and describe the task in two sentences — we will send a range. The agent itself does not quote prices in a chat: a person handles the conversation about price.
The first week is free
For a week the agent works on your real tasks, on the subscription you already have. The hub charges nothing for that week. After it you decide whether to go on and on which model input.
What the cabinet shows
- “Spend” — turns, chats and tokens by day and the cost at the model provider’s published API prices, in dollars, before taxes.
- The breakdown — by agent, by chat and by kind of work: work with people, training, checks and reviews. You can see exactly where the spend went.
- “Subscriptions” — how much each agent used this week on each model input.
How to read the dollar figure: on an API key you pay roughly that much. On a subscription you pay the subscription price, and the figure shows the load the agent puts on it. Training is a one-off cost at the start; there is no need to multiply it by months.
How to tell whether the agent pays off
The price of a token says little about the benefit. It is easier to count from the work:
- List what the agent did during the week: answers to routine questions, briefs, drafts, minutes, reconciliations.
- Estimate how long it would have taken a person, and what an hour of that person costs.
- Compare it with the costs: the share of the subscription or the bill for the key, the infrastructure and the hub’s work for the same period.
Goals with acceptance help here. The owner sets the agent a goal with criteria, and whether it is achieved is decided not by the agent but by an independent reviewer who checks the agent’s report against the criteria. An accepted result is easier to count than “the agent helped”.
What will not be on the bill
- A markup on the model.
- A multiplier on infrastructure.
- A fee per seat: a second agent is not a second bill.
Where to start
Describe in the form the work that comes back again and again. We answer within five working days, and the agent’s first week of work is free.