"How much will it cost?" is usually the first question I get about AI agents, and the answers online are hard to check: wide ranges with no math and no source. I would rather show the math. If you are not sure yet what an agent could do for you, start with AI agent use cases for business, then come back here for the budget.
In this article, I separate what you pay once (the build) from what you pay every month (the model, the tool, the follow-up). For the monthly part, I use the official prices of one model provider, Anthropic, checked on October 17, 2026. The volumes are assumptions: replace them with your own.
In short.
- The AI model is billed by the token (a chunk of a word), sent and generated, at a price per million tokens.
- For a small business this line is often modest: from $0.16 to $34.80 a month in the three cases below (US dollars, before tax).
- The automation tool and hosting come on top: n8n, for example, starts at €20 a month billed annually.
- The biggest items are the setup, then the human review time in the first weeks.
What are the cost items of an AI agent?
An AI agent is a program that receives a request (an email, a message, a form), has an AI model read it along with instructions and documents, then acts: it sorts, replies, drafts something or creates a task. Its cost breaks down into four items.
| Item | When you pay | What drives it |
|---|---|---|
| Setup | Once | Number of tools to connect, quality of your starting documents, share of special cases |
| AI model | Monthly, pay per use | Number of requests, length of texts, model chosen |
| Tool and hosting | Monthly or yearly | Number of executions, tool chosen, cloud or your own server |
| Follow-up and review | Mostly the first weeks | How sensitive the replies are, error rate observed |
Consumer chat assistant subscriptions are not the right benchmark here. An agent connected to your tools goes through the model's programmatic access (the API), billed by usage.
How is the price of an AI model calculated?
Model providers bill by the token: a small piece of text, often part of a word. You pay for two things separately:
- input tokens: everything the agent sends to the model (your instructions, the customer's message, the relevant documents, the conversation history);
- output tokens: what the model writes back. They cost more, five times more with this provider.
Here are the official prices of two Anthropic models, checked on October 17, 2026 on the pricing page of their documentation. Prices are in US dollars, before tax, per million tokens.
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 | $0.01 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
Haiku is the fast, low-cost model, good enough to sort messages or answer from an FAQ. Sonnet is more capable, useful when the agent has to reason over a price list or write a longer text. The same page notes that recent models use a tokenizer that produces about 30% more tokens for the same text: I use that as a safety margin below.
The formula is simple:
monthly cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)
How much does an AI agent cost per month? Three worked cases
I take three agents that are common in a small company. The volumes and text sizes below are assumptions chosen to illustrate the math, not measurements.
| Case | Monthly volume | Input per call | Output per call |
|---|---|---|---|
| Sorting incoming emails | 1,500 emails | 2,500 tokens (instructions + email) | 150 tokens (category + summary) |
| Answering frequent questions | 600 conversations of 4 exchanges, so 2,400 calls | 6,000 tokens (FAQ + history) | 250 tokens |
| Draft quote from a request | 80 requests | 12,000 tokens (price list + request) | 1,500 tokens |
Multiply the volume by the size of each call to get the month's tokens, then apply the prices.
| Case | Tokens per month | With Haiku 5.5 | With Sonnet 5.5 |
|---|---|---|---|
| Sorting emails | 3.75M input, 0.225M output | 0.375 + 0.1125 = $0.49 | 7.50 + 2.25 = $9.75 |
| Frequent questions | 14.4M input, 0.6M output | 1.44 + 0.30 = $1.74 | 28.80 + 6.00 = $34.80 |
| Draft quote | 0.96M input, 0.12M output | 0.096 + 0.06 = $0.16 | 1.92 + 1.20 = $3.12 |
Amounts in US dollars, before tax, calculated on October 17, 2026. With the 30% tokenizer margin, the most expensive case goes from $34.80 to about $45 a month.
How to read this table? For sorting emails, Haiku is usually enough: the task is short and repetitive. For draft quotes, I would rather start with Sonnet, because a pricing mistake costs far more than 3 dollars a month. The frequent-questions case is where the model choice matters most, because every call resends the whole FAQ and the history.
How much does the tool that runs the agent cost?
The model only reads and writes. You need a tool to receive the emails, call the model, save the result in your CRM or spreadsheet and notify someone. I often use n8n for this. Its prices, checked on October 17, 2026 on its official page:
- Starter: €20 a month billed annually, 2,500 workflow executions a month;
- Pro: €50 a month billed annually, 10,000 executions a month;
- self-hosted version: n8n provides a standard version on GitHub that you install on your own server. You then pay for the server and the upkeep time.
An execution is roughly one run through the workflow. With my assumptions, email sorting alone (1,500 executions) fits in Starter, but the three agents together (about 3,980 executions) need Pro. Make and Zapier count differently: I compare all three in n8n vs Make vs Zapier.
Depending on the case, other costs can apply: a business tool that charges for API access, a phone number and minutes for a voice agent (see the AI agent that answers the phone), or a document index if the agent has to search hundreds of files.
How much do setup and maintenance cost?
This is the most variable item, and I won't give a generic price here: two agents that both "answer emails" can take two days or three weeks of work. What moves the price:
- the number of tools to connect (email, CRM, quoting software, calendar);
- the state of your starting documents: an up-to-date FAQ and a clear price list save a lot of time;
- the share of special cases that need separate handling;
- the level of control you want: a draft approved by a person, or direct sending for simple cases.
At Lionea, the price is fixed before work starts: after a free 30-minute call, I build a free mockup on your case, show it to you in a demo, then send the fixed-price quote. The build runs in 2-week sprints, with a working version at the end of each one. After launch, maintenance is a 6-month monthly plan, extendable if needed, and the first month is free. That is how my custom AI agents work.
If you would rather build it yourself, a no-code tool makes it possible: I walk through the steps in how to build an AI agent without code in n8n.
How can you lower the running cost?
- Pick the smallest model that does the job. In the cases above, moving from Sonnet to Haiku cuts the cost by a factor of twenty. Test with the small one first and move up only if quality falls short.
- Cache what repeats. The FAQ sent with every call can be cached: reading it back costs $0.10 per million tokens on Sonnet 5.5 instead of $2. If 4,000 of the 6,000 tokens of each call are read from cache, the frequent-questions case on Sonnet drops from $34.80 to about $16.56, not counting cache writes (billed at $2.50 per million tokens for a 5-minute cache).
- Batch what isn't urgent. Anthropic's pricing page advertises 50% savings with batch processing. Overnight email sorting or a weekly report fit well; a live reply to a customer does not.
- Send only what is needed. A conversation history limited to the last few exchanges and short instructions reduce input tokens.
What are the limits of this estimate?
- Prices change. Model prices move often, down and up. Check the official page again before you decide: my figures date from October 17, 2026.
- Token counts vary. An email with a long signature, an attachment read by the agent or a conversation that drags on push input up. Measure over two real weeks before trusting any estimate.
- The hidden cost is human. At first, someone reviews the agent's replies. That time often matters more than the model bill. It shrinks as instructions and documents improve.
- Currencies and taxes. Model prices are in US dollars before tax; n8n's are in euros. Your final bill depends on exchange rates and the sales tax or VAT that applies to you.
And sometimes you don't need an agent at all. If the task comes up five times a month, or if each reply needs real commercial judgment, a plain automation without AI, or even a well-organized reply template, costs less and breaks less. I say so on the first call when that is the case.
Sources checked on October 17, 2026
- Anthropic, documentation, model pricing page (prices per million tokens, cache, tokenizer)
- Anthropic, pricing page (50% discount for batch processing)
- n8n, pricing page (Starter and Pro plans, included executions, self-hosted version)
Frequently asked questions
How much does an AI agent cost per month?
The AI model cost depends on volume and on the model chosen. In the three cases worked out in this article, it ranges from $0.16 to $34.80 a month before tax, using Anthropic prices checked on October 17, 2026. Add the automation tool and hosting on top.
What is a token in an AI model?
A token is a small piece of text, often part of a word. Providers bill separately for the tokens sent to the model and the tokens it generates, at a price per million tokens.
Is an AI agent more expensive than an AI assistant subscription?
They are different products. A subscription gives one person an assistant. An agent is connected to your tools and billed by usage, plus the tool that runs it. At small volumes, usage often stays low; most of the budget is the setup.
Can you build an AI agent for free?
You can build a first agent with the self-hosted version of n8n, available on GitHub, but you still pay for the server and for the AI model usage. Fully free is rare once the agent handles real work.
How can I lower the cost of an AI agent?
Pick the smallest model that does the job, cache the documents sent with every call, batch what isn't urgent and limit the conversation history sent to the model.