Running 46 AI Agents: An Illustrative Monthly Cost Model
Your AI agent’s token bill may be the smallest thing it costs you. One agent’s month, modelled line by line, shows why.
IN 30 SECONDS
An AI agent’s running cost is everything it takes to keep it working each month, not just the model’s usage bill. In this post’s illustrative model, tokens are about 5% of one agent’s monthly cost and human review about 58%, and I show every assumption.
KEY TAKEAWAYS · 3
- In this model, an agent’s monthly bill has five lines, and the token bill is only the first.
- In this illustrative model, tokens are about 5% of one agent’s monthly cost and human review about 58%. Those shares follow the assumptions, not a rule for every fleet.
- Every dollar figure here is modelled, not an invoice. Redo it with your own numbers.
On this page
The problem: the bill nobody shows you
IN SHORT The cost worth budgeting for comes after launch, and the token bill doesn’t show it.
I deployed about 46 AI agents for internal operations at Integra, where I was CTO until June 2026. People ask what that cost. They mean the token bill, and that’s the least interesting answer I have.
WATCH OUT
Modelled, not measured. The fleet size comes from my Integra work. Every cost below is an illustrative assumption, not an employer or client invoice. Replace the usage, rates and review burden with your own measurements.
The bill people ask about
The bill I actually paid
A model provider’s token bill doesn’t include the review and upkeep work people carry after launch.
Coming up next: So what is on the bill that nobody shows?
Five lines on an agent’s monthly bill
IN SHORT In this model, an agent has five monthly costs, and the token bill is only the first.
WORDS TO KNOW
- Inference
- One call to the AI model: it reads the input and writes an answer.
- Token
- The unit AI providers bill by: a short piece of text, often part of a word.
- Modelled cost
- A cost worked out from assumptions written down in advance, not read off a real invoice.
- Escalation
- The agent handing a case it is not sure about to a person, with what it found.
Each time an agent runs, the model does at least one round of Inference One call to the AI model: it reads the input and writes an answer. See the glossary. The provider counts the text going in and out in Token The unit AI providers bill by: a short piece of text, often part of a word. See the glossary, and bills you for them. That’s the line you see on a dashboard.
STEP BY STEP
Step 1. Inference. Tokens in and out, retries and tool calls land on the provider’s bill.
- Inference Tokens in and out, retries and tool calls land on the provider’s bill.
- Platform Scheduling, storage and monitoring keep the agent running.
- Data upkeep The search index gets rebuilt, and connectors get fixed when a source system changes.
- Human review A person reads the cases the agent hands over and approves its riskiest actions.
- Engineering upkeep An engineer updates the prompts (the model’s instructions) and tests, then moves to new models.
In 2015, Sculley and colleagues at Google wrote in Hidden Technical Debt in Machine Learning Systems (NeurIPS 2015) that “it is common to incur massive ongoing maintenance costs in real-world ML systems.” Agents didn’t change that: each one you add can bring its own upkeep.
ADAM says: An agent is like a car: tokens are the fuel; parking, servicing and a passenger checking risky turns are the other costs. Where it breaks: review cost is a design choice, and agents can share one engine.
Coming up next: So what does one agent’s month look like in dollars?
One agent’s modelled bill, line by line
IN SHORT One read-only assistant, modelled line by line: people reading its doubtful cases cost more than ten times its tokens.
It is an internal assistant that can’t send messages or change records. Each dollar figure is a Modelled cost A cost worked out from assumptions written down in advance, not read off a real invoice. See the glossary. It runs 2,000 times a month, with 8,000 input tokens and 1,000 output tokens per run.
I assume 5% of runs end in an Escalation The agent handing a case it is not sure about to a person, with what it found. See the glossary that needs real review: 3 minutes each, at $60 per reviewer-hour. Engineering upkeep is 2 hours a month at $80 an hour.
Token prices of $1 per million in and $5 per million out are fixed inputs for this illustration, not a current quote. For real rates, check Anthropic’s pricing page, then redo the table.
IN PLAIN WORDS
Out of $521 a month, the model’s tokens are $26. The people reading the cases it hands over are $300.
These shares change with the assumptions. Long contexts, frequent runs or expensive models can make inference the biggest line. They are modelled figures, not costs measured at any employer or client.
ITEM GET
Price the human work first
STAGE CLEAR
What you now know
- The five lines of this monthly cost model, and why tokens are only one.
- Why, in this model, review costs more than ten times the tokens.
Check yourself
Tap or click a question to see the answer.
Illustrative case: after a model switch, the token line of the modelled agent doubles, from $26 to $52 a month. Does the total double too?
No. The total goes from $521 to $547, about 5% more: review and upkeep drive it.
Former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft. Building production AI and regulated Web3 systems.
Have a project to ship or a role to fill?
Discuss your project or roleFollow the engineering notes
Get new articles on production AI and smart contracts by email.