AI agent running costs, worked out for 200 orders a day
What an AI agent costs to run once it is live: token pricing in plain terms, a worked example for email order entry, and three settings that keep the monthly bill small.
This article was generated by AI. Labelled in accordance with Article 50 of the EU AI Act. Responsible for publication: Sophera Consulting.
Ask a vendor about AI agent running costs and you will often hear a monthly platform fee, priced like a software licence. The actual cost of running an agent is something you can work out on a napkin. Anthropic's price list (October 2026) puts Claude Sonnet 5.5 at 2 US dollars per million input tokens and 10 dollars per million output tokens. Those two numbers carry most of the calculation.
What AI agent running costs are made of
Every time an agent handles a case, it sends text to a language model and gets text back. Both directions are billed in tokens, which are fragments of words. Input is everything the model has to read: its instructions, the customer email, the attachment, the relevant slice of your master data. Output is what it writes back, for example a completed order in the format your ERP expects.
For the agents we build, that usage is the only running cost. The model provider bills you directly. There is no operations fee and no subscription. The maintenance agent that watches the system and checks every interface for changes is included in the fixed price. It uses tokens too, and those appear on the same bill from the model provider.
A worked example: order entry from email
Say a wholesaler receives 200 orders a day by email, many with a PDF attached. The agent reads each one, identifies customer, items, quantities and delivery date, checks them against master data and creates the order in the ERP. Anything it cannot match with confidence goes to the sales team with the open question flagged.
For this example we assume 5,000 input tokens and 800 output tokens per order. Your real figures depend on how long your emails and attachments are.
- Input: 200 orders at 5,000 tokens is 1 million tokens a day, or 2 dollars.
- Output: 200 orders at 800 tokens is 160,000 tokens a day, or 1.60 dollars.
- Total: 3.60 dollars a day, about 76 dollars a month over 21 working days.
That comes to just under two cents per order. Set that against the time someone on your team spends retyping orders today. Our article on calculating the cost of manual processes shows how to put a figure on that time.
Three ways to keep the bill small
Use a small model where a small model will do
Deciding whether an email is an order at all does not need the most capable model. Claude Haiku 5.5 costs 0.10 dollars per million input tokens and 0.50 dollars per million output tokens on the same price list, one twentieth of Sonnet. If a small model sorts the inbox and only the actual extraction goes to the larger one, part of the work is billed at the low rate.
Cache the instructions
The instructions the model gets are identical for every order. Providers can cache repeated content like this and charge a fraction for reading it again: 0.10 dollars instead of 2 dollars per million tokens on Sonnet 5.5. If 2,000 of the 5,000 input tokens in our example are fixed instructions, the daily cost drops from 3.60 to roughly 2.85 dollars, as long as orders keep arriving at short intervals.
Run what can wait as an overnight batch
Matching supplier invoices against purchase orders rarely needs an answer within seconds. Jobs like that can go through the batch interface, where Anthropic gives a 50 percent discount on input and output.
When the models run on your own hardware
If you prefer, the models run on your own servers, and the data never leaves the building. For hospitals that is usually the deciding point. That removes the token bill and replaces it with the purchase, power and upkeep of the hardware. A rented dedicated server in a European data centre, at a fixed monthly rate, sits in between. Daily volume decides which option pays off, and we work that out for the whole organisation before anyone commits, not for a single process.
What an honest quote shows
You can calculate an agent's running costs before signing anything. Ask for that calculation using your own volumes: cases per day, tokens per case, model and price. If the quote shows a flat monthly fee for "operations" or "platform" instead, you are paying for more than compute, and it is fair to ask what for.
Sophera Consulting builds the agent at a fixed price in one to two days, with a maintenance agent and no subscription. After that, you pay only for what the models actually use. Which process to start with, and what volumes to plan for, is what we work out in the free automation check.
This article was created with the help of AI.