AI Agent Cost Calculator

An AI agent has two costs: a one-time build and a monthly running cost. Describe the job and you get both, with the math shown, on usage-based billing and on a flat subscription.

The job

Inbox, CRM, calendar, spreadsheet. Count each one.

The whole thread, ticket or form it has to read.

A reply draft, a summary or a record update.

Background it needs every time

What the agent has to know about your business before it can do the job.

The agent drafts or proposes, and nothing goes out until someone says yes.

How it is billed and where it runs

A small model handles sorting and short replies. Pick a larger one for judgment calls.

Plan, for the subscription route
Where the agent runs

Printed on the saved result.

AI agent cost estimate

Your agent cost estimate

Monthly running cost, usage-based

$19.41to$31.51

Claude Sonnet 5.5 for 500 emails a month, hosting included. The lower figure uses prompt caching, the higher one does not.

On Claude Max instead

$106 a month

Cost per email

$0.027 to $0.051

Model calls per task

3

Your review time

8.3 hours a month

Usage-based billing costs less than Claude Max for this job, even without prompt caching.

Usage-based route: the token math
ItemAmount
Text the model reads23,367 tokens a task, 11,683,500 a month, at $2 per million$23.37
Text the model writes427 tokens a task, 213,500 a month, at $10 per million$2.14
HostingA small cloud server$6.00
Monthly total, no caching$31.51
Subscription route
ItemAmount
Claude MaxFlat price. Usage limits reset every five hours and every week.$100
HostingA small cloud server$6.00
Monthly total$106

One-time build cost, agency-built

$8,000to$40,000

Published agency ranges for this kind of agent: one job, one or two connected tools. Building it yourself on a subscription has no build fee and costs your own time.

Where the build range comes from
Source and categoryRange
Riseup LabsReactive or simple reflex agent$8,000 to $20,000
RaftLabsSingle-workflow agent$15,000 to $40,000
Upkeep each year, if an agency maintains it15% to 30% of the build cost, per Riseup Labs$1,200 to $12,000

One job, 2 tools and a person approving the work is small enough for a fixed-price setup. See what the first agent setup includes.

How we calculated this

Words to tokens.Models bill by the token, a piece of a word. Google's rule of thumb is that 100 tokens is about 60 to 80 English words, and this calculator uses 70. Anthropic says its current models produce about 30% more tokens for the same text, so that is added for Claude Sonnet 5.5.

Calls per task. An agent does not answer in one go. It calls the model once to start and once more after each tool it uses, and every call resends everything so far. That is 3 calls here. Each call carries 6,415 tokens of instructions, background (3,000 words) and tool descriptions, plus what the task itself adds.

Model cost. Tokens read times the input price, plus tokens written times the output price, times your tasks a month. Claude Sonnet 5.5 is listed at $2 per million tokens read and $10 per million written, and $0.10 per million for text it has already read in the same task (prompt caching). Source: Anthropic, Claude API pricing, checked 2026-11-17.

Subscription and hosting. Plan prices are from Claude plans and pricing, checked 2026-11-17. Plans cap usage and do not publish the cap as a number of tasks, so the usage-based figure is the best gauge of how heavy the job is: the further it sits above the plan price, the more likely the plan runs out. The cloud server price is from DigitalOcean, Droplet pricing, checked 2026-11-17.

Build cost. The range runs from the lowest to the highest figure the two agency guides in the table publish for the closest category. A job with more than 2 tools is matched to their multi-tool categories, and customer questions to their support categories. Prices checked 2026-11-17.

What is assumed, not measured. Each connected tool is used once per task, takes about 150 words to describe to the model, 40 words to request and returns about 300words. These are round planning numbers. Retries, long conversations and web searches cost extra, and caching between tasks can save more. A week of real runs gives you exact token counts on the provider's bill.

Save your result

Example: an email triage agent for a small business

An AI agent that triages 500 emails a month, reads about 400 words and drafts about 150 words for each, knows a 3,000-word handbook about the business and connects to 2 tools costs $19.41 to $31.51 a month to run on usage-based billing with Claude Sonnet 5.5, including a $6 cloud server. That is $0.027 to $0.051 per email, at $2 per million tokens read and $10 per million written (prices checked 2026-11-17).

The same agent on a Claude Max subscription costs $106 a month. Agencies publish $8,000 to $40,000 to build an agent of this size, and a person approving every draft spends about 8.3 hours a month on it.

Example agent cost estimate

Monthly running cost, usage-based

$19.41to$31.51

Claude Sonnet 5.5 for 500 emails a month, hosting included. The lower figure uses prompt caching, the higher one does not.

On Claude Max instead

$106 a month

Cost per email

$0.027 to $0.051

Model calls per task

3

Your review time

8.3 hours a month

Usage-based billing costs less than Claude Max for this job, even without prompt caching.

Usage-based route: the token math
ItemAmount
Text the model reads23,367 tokens a task, 11,683,500 a month, at $2 per million$23.37
Text the model writes427 tokens a task, 213,500 a month, at $10 per million$2.14
HostingA small cloud server$6.00
Monthly total, no caching$31.51
Subscription route
ItemAmount
Claude MaxFlat price. Usage limits reset every five hours and every week.$100
HostingA small cloud server$6.00
Monthly total$106

One-time build cost, agency-built

$8,000to$40,000

Published agency ranges for this kind of agent: one job, one or two connected tools. Building it yourself on a subscription has no build fee and costs your own time.

Where the build range comes from
Source and categoryRange
Riseup LabsReactive or simple reflex agent$8,000 to $20,000
RaftLabsSingle-workflow agent$15,000 to $40,000
Upkeep each year, if an agency maintains it15% to 30% of the build cost, per Riseup Labs$1,200 to $12,000

One job, 2 tools and a person approving the work is small enough for a fixed-price setup. See what the first agent setup includes.

How we calculated this

Words to tokens.Models bill by the token, a piece of a word. Google's rule of thumb is that 100 tokens is about 60 to 80 English words, and this calculator uses 70. Anthropic says its current models produce about 30% more tokens for the same text, so that is added for Claude Sonnet 5.5.

Calls per task. An agent does not answer in one go. It calls the model once to start and once more after each tool it uses, and every call resends everything so far. That is 3 calls here. Each call carries 6,415 tokens of instructions, background (3,000 words) and tool descriptions, plus what the task itself adds.

Model cost. Tokens read times the input price, plus tokens written times the output price, times your tasks a month. Claude Sonnet 5.5 is listed at $2 per million tokens read and $10 per million written, and $0.10 per million for text it has already read in the same task (prompt caching). Source: Anthropic, Claude API pricing, checked 2026-11-17.

Subscription and hosting. Plan prices are from Claude plans and pricing, checked 2026-11-17. Plans cap usage and do not publish the cap as a number of tasks, so the usage-based figure is the best gauge of how heavy the job is: the further it sits above the plan price, the more likely the plan runs out. The cloud server price is from DigitalOcean, Droplet pricing, checked 2026-11-17.

Build cost. The range runs from the lowest to the highest figure the two agency guides in the table publish for the closest category. A job with more than 2 tools is matched to their multi-tool categories, and customer questions to their support categories. Prices checked 2026-11-17.

What is assumed, not measured. Each connected tool is used once per task, takes about 150 words to describe to the model, 40 words to request and returns about 300words. These are round planning numbers. Retries, long conversations and web searches cost extra, and caching between tasks can save more. A week of real runs gives you exact token counts on the provider's bill.

What an AI agent costs

An AI agent costs a one-time amount to build and a monthly amount to run, and the two are very different sizes. For a small business agent that does one job, running it usually costs tens of dollars a month, while agencies publish build prices in the thousands. Most guides blend the two into one large number. The calculator keeps them apart so you can see which one you are really deciding about.

How the monthly running cost is worked out

Model providers bill by the token, which is a piece of a word. You pay one price for every token the model reads and a higher price for every token it writes. The monthly cost is tokens read times the input price, plus tokens written times the output price, times the number of tasks.

What surprises people is how much an agent reads. It does not handle a task in one go. It calls the model once to start and once more after each tool it uses, and every call resends the instructions, the background about your business and everything gathered so far. A 400-word email can turn into more than 20,000 tokens read by the time the draft is written.

Three things move the number most:

  • The model. A small model costs a twentieth of a mid-size one at current list prices.
  • The background. Whatever the agent has to know about your business is resent on every call, so a long reference costs more than a one-page brief.
  • The number of tools. Each tool adds a model call and makes every later call longer.

Prompt caching brings the cost down. Providers charge a fraction of the input price for text the model has already read, which is why the calculator shows a range.

Subscription or usage-based billing

There are two ways to pay for the model. A subscription is a flat monthly price with usage limits. Usage-based billing through the provider's API charges for exactly the tokens used, with no limit.

We run our own operations agent on a subscription. It runs on Claude Code on a Mac mini in the office, and the only new costs are the Claude subscription and Slack. The full setup is in how we built a Slack agent that runs our operations.

That choice has a price. The Mac mini lost power for about two days in late August, nothing ran, and nothing reported it. If an agent is critical to your business, host it in the cloud and run it on usage-based billing instead. It costs more, but it does not go dark when the power goes out.

What the build cost covers

The build is mostly not the model. It is writing down how the job is done, connecting each tool with access limited to what the job needs, adding an approval step, and testing on real work until the output can be trusted. Published agency ranges are wide because the scope is: one job with two tools is a different project from an agent that works across your whole business. Multi-agent systems sit above everything in the calculator, at $120,000 to $400,000 in the same two guides.

The cheapest way to keep the build small is to start with one narrow job. That is also what makes an agent useful, for the reasons in what we learned building agents that actually work.

What the estimate leaves out

The estimate covers model usage, hosting and the build. It does not price the tools the agent connects to, because you are likely paying for your inbox, CRM and phone system already. It does not count retries, web searches or very long conversations, which add to the model bill. Token counts are estimated from word counts, so treat the result as a planning figure and check it against the provider's bill after a week of real runs.

Common Questions

How much does an AI agent cost?

An AI agent has a one-time build cost and a monthly running cost. For a small single-job agent, agencies publish build prices from $8,000 to $40,000, and building it yourself costs your time instead. Running it is far cheaper: the email triage example on this page comes to about $19 to $32 a month on usage-based billing, or $106 a month on a flat subscription.

How much does it cost to build an AI agent?

Published agency prices start at $8,000 to $40,000 for an agent that does one job with one or two tools, and reach $45,000 to $120,000 for one that works across several tools. Those figures come from Riseup Labs and RaftLabs. You can also build a first agent yourself on a subscription, which has no build fee and takes your own time to set up and tune.

How much does an AI agent cost per month?

The monthly cost is the model usage plus hosting. Model usage is the tokens the agent reads and writes per task, times the model's price, times your tasks a month. In the example on this page, 500 emails a month on Claude Sonnet 5.5 cost about $13 to $26 in model usage plus $6 for a small cloud server. If an agency maintains the agent, Riseup Labs puts upkeep at 15% to 30% of the build cost per year.

Is a subscription or usage-based billing cheaper for an AI agent?

Usage-based billing is cheaper at low volume and a flat subscription is cheaper at high volume, as long as the work fits inside the plan's usage limits. In the example, 500 emails a month cost about $32 on usage against $106 on Claude Max, while 5,000 emails a month cost about $261 on usage against the same $106. Plans do not publish their limits as a number of tasks, and an agent that answers customers directly belongs on usage-based billing.

Why does an AI agent cost more to run than a chatbot?

An agent calls the model several times for one task. It calls once to start and again after each tool it uses, and every call resends the instructions, the background and everything gathered so far. Three calls on a 400-word email means the model reads about 23,000 tokens, not the few hundred in the email itself.

What is the cheapest model for an AI agent?

The small models are the cheapest. Claude Haiku 5.5 and gpt-6-luna are both listed at $0.10 per million tokens read and $0.50 per million written, which is 20 times less than Claude Sonnet 5.5 at $2 and $10. A small model is usually enough for sorting, tagging and short replies. Use a larger one where the agent has to make judgment calls.

What other costs come with an AI agent?

Plan for the time a person spends approving the agent's work, the tools it connects to (most of which you already pay for), and upkeep when your business changes. For a small agent, token cost is often the smallest of these. At one minute per email, approving 500 drafts a month is about 8 hours of someone's time.

Two ways to get started

First Agent Setup

Get your first agent running on one job

We set up one AI agent for the job you just priced: one job, up to two of the tools you already use, and nothing goes out until you approve it. It runs on your own Claude plan or API account, so the running cost in your estimate is billed to you by the provider and is not part of this price.

  • A kickoff call to walk through the job, using your estimate from the calculator
  • A written playbook for the job: what the agent does, when it runs and what needs your approval
  • Connections to up to two of your tools, with access limited to what the job needs
  • An approval step in Slack or email, so the agent drafts and proposes and you decide
  • Test runs on your real work, a handoff walkthrough and 14 days of fixes after it starts
$1,500one time

Turnaround: Running within 10 business days of your kickoff call.

You pay on Stripe's secure checkout page. We never see your card number.

Build Plan and Fixed Quote

Get the plan and one fixed price before you pay to build

We look at how the work runs today and tell you plainly what is worth building, what a ready-made product or a simple automation would cover, where an AI agent would help and what to leave alone. You get it in writing, with a fixed price for the first build, so you know the full number before any work begins. The plan is yours to keep, whether we build it, someone else does, or nobody does.

  • A short questionnaire about how the work runs today, reviewed before we talk. If you already wrote a scope of work, we start from it
  • A one-hour call that walks through the work you want off your plate
  • A written plan that says for each part: build it, buy a ready-made product, automate it, hand it to an AI agent, or leave it alone
  • What the recommended setup would cost to build and to run every month
  • A fixed-price proposal for the first build, with scope, timeline and deliverables, valid for 30 days
$495one time

Turnaround: Written plan within 5 business days of your call.

You pay on Stripe's secure checkout page. We never see your card number.